Do We Need More Training Data or Better Models for Object Detection?
Abstract
(Work performed while at UC Irvine) Datasets for training object recognition systems are steadily growing in size. This paper investigates the question of whether existing detectors will continue to improve as data grows, or if models are close to saturating due to limited model complexity and the Bayes risk associated with the feature spaces in which they operate. We focus on the popular paradigm of scanning-window templates defined on oriented gradient fea-tures, trained with discriminative classifiers. We investigate the performance of mixtures of templates as a function of the number of templates (complexity) and the amount of training data. We find that additional data does help, but only with correct regularization and treatment of noisy examples or “outliers ” in the training data. Surprisingly, the per-formance of problem domain-agnostic mixture models appears to saturate quickly (∼10 templates and ∼100 positive training examples per template). However, compositional mixtures (implemented via composed parts) give much better performance because they share parameters among templates, and can synthesize new templates not encountered during training. This suggests there is still room to improve performance with linear classifiers and the existing feature space by improved representations and learning algo-rithms. 1
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
Cited by
Sorted by citations
Microsoft COCO: Common Objects in Context
Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., & Zitnick, C. L. (2014) · Lecture Notes in Computer Science · doi:10.1007/978-3-319-10602-1_48 · Cited by 45,460 · see citing works
“Consistent with past observations [46], we find that including difficult (non-iconic) images during training may not always help.”
BackgroundSelf-citationRich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation
Girshick, R., Donahue, J., Darrell, T., & Malik, J. (2014) · 2014 IEEE Conference on Computer Vision and Pattern Recognition · doi:10.1109/cvpr.2014.81 · Cited by 32,684 · see citing works
“Additionally, [42] shows that even when more data is available, DPM does not easily benefit from it.”
BackgroundGirshick, R. (2015) · 2015 IEEE International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2015.169 · Cited by 28,105 · see citing works
Shin, H.-C., Roth, H. R., Gao, M., Lu, L., Xu, Z., Nogues, I., Yao, J., Mollura, D., & Summers, R. M. (2016) · IEEE Transactions on Medical Imaging · doi:10.1109/tmi.2016.2528162 · Cited by 5,759 · see citing works
Microsoft COCO: Common Objects in Context
Lin, T.-Y., Maire, M., Belongie, S., Bourdev, L., Girshick, R., Hays, J., Perona, P., Ramanan, D., Zitnick, C. L., & Dollár, P. (2014) · arXiv · doi:10.48550/arxiv.1405.0312 · Cited by 2,579 · see citing works
Self-citationFast R-CNNPreprint
Girshick, R. B. (2015) · arXiv (Cornell University) · doi:10.48550/arxiv.1504.08083 · Cited by 29,416 · see citing works
Tomislav Hengl, A., Nussbaum, M., Wright, M. N., Heuvelink, G. B. M., & Gräler, B. (2018) · PeerJ · doi:10.7717/peerj.5518 · Cited by 1,130 · see citing works
“It does not have an718 explicit (geo)statistical model as a starting point, so that no mathematical derivations are possible at all.719 Also, just adding more and more data to the system does not necessarily mean that the predictions will720 automatically become better (Zhu et al., 2012).”
BackgroundChen, K., Chen, H., Zhou, C., Huang, Y., Qi, X., Shen, R., Liu, F., Zuo, M., Zou, X., Wang, J., et al. (2019) · Water Research · doi:10.1016/j.watres.2019.115454 · Cited by 606 · see citing works
Object Detection Networks on Convolutional Feature Maps
Ren, S., He, K., Girshick, R., Zhang, X., & Sun, J. (2017) · IEEE Transactions on Pattern Analysis and Machine Intelligence · doi:10.1109/tpami.2016.2601099 · Cited by 454 · see citing works
“PM+CNN methods [24 ,11 29] and R-CNN/SPPnet. The DPM model is a less powerful classifier than a multi-layer network, whether applied to deep convolutional features or shallow HOG features. The work in [34] shows that a DPM classifier has diminishing improvement when given more training data. To method trained on mAP are…”
MethodFrom generic to specific deep representations for visual recognition
Azizpour, H., Razavian, A. S., Sullivan, J., Maki, A., & Carlsson, S. (2015) · IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) · doi:10.1109/cvprw.2015.7301270 · Cited by 401 · see citing works
“Zhu et al.[48] suggest that increasing data is less effective than increasing the complexity of models or richness of representation and the former is prone to early performance saturation.”
BackgroundSliding Shapes for 3D Object Detection in Depth Images
Song, S., & Xiao, J. (2014) · Lecture notes in computer science · doi:10.1007/978-3-319-10599-4_41 · Cited by 361 · see citing works
“[28–30] showed that the latter option is much simpler and has the generalizability to all object categories.”
BackgroundA survey on data‐efficient algorithms in big data era
Adadi, A. (2021) · Journal of Big Data · doi:10.1186/s40537-021-00419-9 · Cited by 331 · see citing works
“[63] highlighted by the evidence a surprising observation that is at some point,”
BackgroundGlioma Grading on Conventional MR Images: A Deep Learning Study With Transfer Learning
Yang, Y., Yan, L., Zhang, X., Han, Y., Nan, H., Hu, Y., Hu, B., Yan, S., Zhang, J., Cheng, D.-L., et al. (2018) · Frontiers in Neuroscience · doi:10.3389/fnins.2018.00804 · Cited by 317 · see citing works
“Thus, it is crucial to find the trade-off between suitable model and data size (Zhu et al., 2012).”
BackgroundFactors of Transferability for a Generic ConvNet Representation
Azizpour, H., Razavian, A. S., Sullivan, J., Maki, A., & Carlsson, S. (2016) · IEEE Transactions on Pattern Analysis and Machine Intelligence · doi:10.1109/tpami.2015.2500224 · Cited by 301 · see citing works
“[42] suggest that increasing the amount of training data is less effective than increasing the complexity or richness of the representation or the classification models used and that the former is prone to early performance saturation.”
BackgroundLearning Everything about Anything: Webly-Supervised Visual Concept Learning
Divvala, S. K., Farhadi, A., & Guestrin, C. (2014) · 2014 IEEE Conference on Computer Vision and Pattern Recognition · doi:10.1109/cvpr.2014.412 · Cited by 267 · see citing works
“These annotations can only tackle simple appearance variations of an object [51].”
BackgroundHuman Pose Estimation Using Body Parts Dependent Joint Regressors
Dantone, M., Gall, J., Leistner, C., & Gool, L. V. (2013) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr.2013.391 · Cited by 293 · see citing works
“While this approach outperforms classical pictorial structure models for human pose estimation, it has been shown in [41] that the used templates, which are scanningwindow templates trained with linear SVMs on HOG features [7], are very sensitive to noise and limit the performance.”
MethodHOGgles: Visualizing Object Detection Features
Vondrick, C., Khosla, A., Malisiewicz, T., & Torralba, A. (2013) · 2013 IEEE International Conference on Computer Vision · doi:10.1109/iccv.2013.8 · Cited by 282 · see citing works
“[24] try to determine whether we have reached Bayes risk for HOG.”
BackgroundSelf-citationSeeking the Strongest Rigid Detector
Benenson, R., Mathias, M., Tuytelaars, T., & Gool, L. V. (2013) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr.2013.470 · Cited by 257 · see citing works
“It is well-known that the data used to train an algorithm is just as important as the algorithm itself [18].”
BackgroundHistograms of Sparse Codes for Object Detection
Ren, X., & Ramanan, D. (2013) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr.2013.417 · Cited by 239 · see citing works
“There are evidences that local features are most crucial for detection [23], and we may already be saturating the capacity of HOG [36].”
BackgroundSelf-citationKnowledge-Guided Deep Fractal Neural Networks for Human Pose Estimation
Ning, G., Zhang, Z., & He, Z. (2017) · IEEE Transactions on Multimedia · doi:10.1109/tmm.2017.2762010 · Cited by 180 · see citing works
Depth-Based Hand Pose Estimation: Data, Methods, and Challenges
Supančič, J. S., Rogez, G., Yang, Y., Shotton, J., & Ramanan, D. (2015) · IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2015.217 · Cited by 172 · see citing works
Self-citationCapturing Long-Tail Distributions of Object Subcategories
Zhu, X., Anguelov, D., & Ramanan, D. (2014) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr.2014.122 · Cited by 135 · see citing works
“Subcategory discovery: Estimating subcategories is surprisingly hard; clustering based on keypoints [16, 5] and appearance [11, 3] have provided only modest performance increases [34].”
BackgroundSelf-citationChallenges of Hate Speech Detection in Social Media
Kovács, G., Alonso, P., & Saini, R. (2021) · SN Computer Science · doi:10.1007/s42979-021-00457-3 · Cited by 162 · see citing works
“One could argue that the benefits of more data are already well established, but that is not always the case [27, 100].”
BackgroundDiagnosis of depression by behavioural signals
Cummins, N., Joshi, J., Dhall, A., Sethu, V., Goecke, R., & Epps, J. (2013) · doi:10.1145/2512530.2512535 · Cited by 138 · see citing works
“We argue that in certain cases more training data does not always guarantee better system performance [49].”
A Performance Evaluation of Machine Learning-Based Streaming Spam Tweets Detection
Chen, C., Zhang, J., Xie, Y., Xiang, Y., Zhou, W., Hassan, M. M., Alelaiwi, A. A., & Alrubaian, M. (2015) · IEEE Transactions on Computational Social Systems · doi:10.1109/tcss.2016.2516039 · Cited by 131 · see citing works
“More preprocesses, such as developing more discriminant features or cleaning training data [36], should be done to further improve the performance.”
BackgroundLearning to Detect Vehicles by Clustering Appearance Patterns
Ohn-Bar, E., & Trivedi, M. M. (2015) · IEEE Transactions on Intelligent Transportation Systems · doi:10.1109/tits.2015.2409889 · Cited by 120 · see citing works
“Unfortunately, learning models even just for 5 scales (as in [8]) is very costly, especially due to having to learn multiple components per scale.”
BackgroundExpanded Parts Model for Human Attribute and Action Recognition in Still Images
Sharma, G., Jurie, F., & Schmid, C. (2013) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr.2013.90 · Cited by 104 · see citing works
“This is in line with a recent study [40], in “mixture of template’ like formulations, there are clean interpretable templates along with noisy templates which correspond to background.”
BackgroundResultFeng, Z., Hu, G., Kittler, J., Christmas, W. J., & Wu, X. (2015) · IEEE Transactions on Image Processing · doi:10.1109/tip.2015.2446944 · Cited by 89 · see citing works
“In contrast, a discriminative model directly predicts the shape update, using a mapping function from a feature space to the shape space, to guide the evolution from an initial shape estimate to the true positions.”
BackgroundTransportation mode detection – an in-depth review of applicability and reliability
Prelipcean, A. C., Gidófalvi, G., & Susilo, Y. O. (2016) · Transport Reviews · doi:10.1080/01441647.2016.1246489 · Cited by 97 · see citing works
“…learning methods has drawbacks, most notable the dependance on large amounts of labeled data needed to train classifiers (Banko & Brill, 2001; Zhu, Vondrick, Ramanan, & Fowlkes, 2012), the difficulty of dealing with imbalanced datasets (Van Hulse, Khoshgof-taar, & Napolitano, 2007), and the…”
BackgroundSubcategory-Aware Object Classification
Dong, J., Xia, W., Chen, Q. K., Feng, J., Huang, Z., & Yan, S. (2013) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr.2013.112 · Cited by 93 · see citing works
“It has been shown that models trained by “clean” subsets of images usually perform better than trained with all images [39].”
BackgroundDetecting Avocados to Zucchinis: What Have We Done, and Where Are We Going?
Russakovsky, O., Deng, J., Huang, Z., Berg, A. C., & Fei-Fei, L. (2013) · IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2013.258 · Cited by 84 · see citing works
“[25] analyzed the relative impact of adding more training data versus building better detection models.”
BackgroundUnderstanding Objects in Detail with Fine-Grained Attributes
Vedaldi, A., Mahendran, S., Tsogkas, S., Maji, S., Girshick, R., Kannala, J., Rahtu, E., Kokkinos, I., Blaschko, M. B., Weiss, D. A., et al. (2014) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr.2014.463 · Cited by 86 · see citing works
“3.2 also shows that, contrary to what has been observed in other object detection tasks in the literature [45], detection of object parts can be improved by adding a significant number of part templates to a state-of-the-art detector.”
InfluentialDepth-Based Hand Pose Estimation: Methods, Data, and Challenges
Supančič, J. S., Rogez, G., Yang, Y., Shotton, J., & Ramanan, D. (2018) · International Journal of Computer Vision · doi:10.1007/s11263-018-1081-7 · Cited by 82 · see citing works
“We specifically train an exemplar part model (EPM) constrained to model deformations consistent with 3D exemplars [43], which will be described further in a tech report.”
InfluentialSelf-citationAutomated Face Analysis for Affective Computing
Calvo, R. A., D’Mello, S. K., Gratch, J. S., Kappas, A., Cohn, J. F., & la Torre, F. D. (2014) · Oxford University Press eBooks · doi:10.1093/oxfordhb/9780199942237.013.020 · Cited by 82 · see citing works
“Initial evidence from object recognition (X. Zhu et al., 2012), gesture recognition (Sutton, 2011), and smile detection (Whitehill et al., 2009) suggest that datasets orders of magnitude larger than those available to date will be needed to achieve optimal AFA.”
BackgroundAl-Sahaf, H., Zhang, M., Al-Sahaf, A., & Johnston, M. (2017) · IEEE Transactions on Evolutionary Computation · doi:10.1109/tevc.2017.2685639 · Cited by 74 · see citing works
“[43] investigated the tradeoff between”
Object detection through search with a foveated visual system
Akbaş, E., & Eckstein, M. P. (2017) · PLoS Computational Biology · doi:10.1371/journal.pcbi.1005743 · Cited by 76 · see citing works
“respectively. Although there are efforts to go beyond these standard choices (e.g. new features [37], [46]; alternative detection models [22], [46], whether object parts should be modeled or not [8], [54]; and alternative search methods [13], [14], [20], [24], [46]), HOG, mixture of linear templates and SW form the cr…”
BackgroundEffects of target typicality on categorical search
Maxfield, J. T., Stalder, W. D., & Zelinsky, G. J. (2014) · Journal of Vision · doi:10.1167/14.12.1 · Cited by 76 · see citing works
Body Parts Dependent Joint Regressors for Human Pose Estimation in Still Images
Dantone, M., Gall, J., Leistner, C., & Van Gool, L. (2014) · IEEE Transactions on Pattern Analysis and Machine Intelligence · doi:10.1109/tpami.2014.2318702 · Cited by 73 · see citing works
“Index Terms—Human pose estimation, fashion, random forest, regression, classification F”
Learning Collections of Part Models for Object Recognition
Endres, I., Shih, K. J., Jiaa, J., et al. (2013) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr.2013.126 · Cited by 73 · see citing works
“[22] concludes that finding better methods to organize examples and parts into visual sub-categories is the most promising direction for future research.”
Williams, D. P. (2020) · IEEE Journal of Oceanic Engineering · doi:10.1109/joe.2019.2963041 · Cited by 74 · see citing works
“eventually hit a performance plateau: beyond a certain point, incorporating more data ceases to improve performance [15].”
BackgroundVideo text detection and recognition: Dataset and benchmark
Nguyen, P. X., Wang, K., & Belongie, S. (2014) · IEEE Winter Conference on Applications of Computer Vision · doi:10.1109/wacv.2014.6836024 · Cited by 71 · see citing works
“[24] suggest that we can grow the model complexity by adding mixture models to capture the “sub-category” structure of the data.”
MethodA Constructive Prediction of the Generalization Error Across Scales
Rosenfeld, J. S., Rosenfeld, A., Belinkov, Y., & Shavit, N. (2019) · arXiv · doi:10.48550/arxiv.1909.12673 · Cited by 74 · see citing works
“Data scaling: It has long been recognized that more data improves performance, and various studies report such trends in both computer vision (e.g., Zhu et al., 2012; Sun et al., 2017) and language processing tasks (e.g., Banko & Brill, 2001; Talmor & Berant, 2019).”
BackgroundData-Driven Design of Intelligent Wireless Networks: An Overview and Tutorial
Kulin, M., Fortuna, C., Poorter, E. D., Deschrijver, D., & Moerman, I. (2016) · Sensors · doi:10.3390/s16060790 · Cited by 63 · see citing works
“In some cases, this may mean that the models where more training/testing data is available might capture the underlying process better than others [140].”
BackgroundInfrared Thermal Image-Based Sustainable Fault Detection for Electrical Facilities
Kim, J. S., Choi, K. N., & Kang, S. W. (2021) · Sustainability · doi:10.3390/su13020557 · Cited by 60 · see citing works
Ghiasi, G., Yang, Y., Ramanan, D., & Fowlkes, C. C. (2014) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr.2014.308 · Cited by 57 · see citing works
“Generating a large number of quality clusters is a surprisingly hard problem; typical approaches of clustering image patches [6], clustering keypoint annotation [2], or even manual grouping [27] have shown only modest performance increases with increasing numbers of clusters.”
BackgroundSelf-citationRegularized Max Pooling for Image Categorization
Hoai, M. (2014) · British Machine Vision Conference (BMVC) · doi:10.5244/c.28.32 · Cited by 58 · see citing works
“If the locations of parts are unknown, the performance significantly drops [32].”
BackgroundFacial Action Unit Detection Using Active Learning and an Efficient Non-linear Kernel Approximation
Sénéchal, T., McDuff, D., & el Kaliouby, R. (2015) · IEEE International Conference on Computer Vision Workshops · doi:10.1109/iccvw.2015.11 · Cited by 58 · see citing works
Learning a Part-Based Pedestrian Detector in a Virtual World
Xu, J., Vázquez, D., López, A. M., et al. (2014) · IEEE Transactions on Intelligent Transportation Systems · doi:10.1109/tits.2014.2310138 · Cited by 56 · see citing works
“However, part sharing has recently shown benefits in tasks such as object detection and pose estimation [7], [14]–[16].”
BackgroundMethodInfluentialObject detection and tracking on UAV RGB videos for early extraction of grape phenotypic traits
Ariza-Sentís, M., Baja, H., Vélez, S., & Valente, J. R. P. (2023) · Computers and Electronics in Agriculture · doi:10.1016/j.compag.2023.108051 · Cited by 55 · see citing works
iSurveillance: Intelligent framework for multiple events detection in surveillance videos
Lim, M. K., Tang, S., & Chan, C. S. (2014) · Expert Systems with Applications · doi:10.1016/j.eswa.2014.02.003 · Cited by 55 · see citing works
Analysis by Synthesis: 3D Object Recognition by Object Reconstruction
Hejrati, M., & Ramanan, D. (2014) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr.2014.314 · Cited by 52 · see citing works
“Zhu et al. [23] point out that DPMs implicitly synthesize a set of deformed templates by searching over possible latent values.”
BackgroundSelf-citationLearning And-Or Model to Represent Context and Occlusion for Car Detection and Viewpoint Estimation
Wu, T., Li, B. H., & Zhu, S. C. (2015) · IEEE Transactions on Pattern Analysis and Machine Intelligence · doi:10.1109/tpami.2015.2497699 · Cited by 51 · see citing works
Cho, Y., Bianchi‐Berthouze, N., Marquardt, N., & Julier, S. (2018) · CHI Conference on Human Factors in Computing Systems · doi:10.1145/3173574.3173576 · Cited by 50 · see citing works
“Such techniques make it easier to extend the recognition capabilities to new types of materials by simply adding sufficient new training data [18,57].”
BackgroundWhat Is Holding Back Convnets for Detection?
Pepik, B., Benenson, R., Ritschel, T., & Schiele, B. (2015) · Lecture notes in computer science · doi:10.1007/978-3-319-24947-6_43 · Cited by 45 · see citing works
“Most of these works focused on DPM-like detectors, which can only make limited use of large training sets [41].”
BackgroundBuilding Part-Based Object Detectors via 3D Geometry
Shrivastava, A., & Gupta, A. (2013) · IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2013.219 · Cited by 39 · see citing works
Impact of compressed and down-scaled training images on vehicle detection in remote sensing imagery
Karim, S., Zhang, Y., Yin, S., Laghari, A. A., & Brohi, A. A. (2019) · Multimedia Tools and Applications · doi:10.1007/s11042-019-08033-x · Cited by 42 · see citing works
“A detailed analysis has consequently assumed that the optimization of extracted features is advantageous for detection algorithms [75].”
BackgroundVisualizing Object Detection Features
Vondrick, C., Khosla, A., Pirsiavash, H., Malisiewicz, T., & Torralba, A. B. (2016) · International Journal of Computer Vision · doi:10.1007/s11263-016-0884-7 · Cited by 38 · see citing works
“By increasing γ , the inversions will look more different, at the expense of matching less in feature space.”
BackgroundSelf-citationXu, Q., & Zheng, R. (2017) · doi:10.1109/infocom.2017.8057034 · Cited by 40 · see citing works
“More annotated instances generally imply better classifier performance [4], [5].”
BackgroundRole of training data variability on classifier performance and generalizability
Therrien, R., & Doyle, S. T. (2018) · SPIE Medical Imaging: Digital Pathology · doi:10.1117/12.2293919 · Cited by 39 · see citing works
“In deep learning classification of natural images (those which contain everyday objects such as people, cars, buildings and pets), it has been demonstrated that large training dataset sizes are required to fully saturate training and achieve high performance [12], [13].”
BackgroundA sematic and prior-knowledge-aided monocular localization method for construction-related entities
Fang, Q., Li, H., Luo, X., et al. (2020) · Computer-Aided Civil and Infrastructure Engineering · doi:10.1111/mice.12541 · Cited by 38 · see citing works
Are all training examples equally valuablePreprint
Lapedriza, À., Pirsiavash, H., Bylinskii, Z., & Torralba, A. B. (2013) · arXiv (Cornell University) · doi:10.13016/m2jp7z-giib · Cited by 53 · see citing works
How much training data for facial action unit detection?
Girard, J. M., Cohn, J. F., Jeni, L. A., Lucey, S., & la Torre, F. D. (2015) · International Conference on Automatic Face and Gesture Recognition · doi:10.1109/fg.2015.7163106 · Cited by 37 · see citing works
“In a study on object detection, Zhu et al. [42] found the counter-intuitive result that larger training sets sometimes led to worse classification performance than smaller training sets.”
BackgroundExpanded Parts Model for Semantic Description of Humans in Still Images
Sharma, G., Jurie, F., & Schmid, C. (2016) · IEEE Transactions on Pattern Analysis and Machine Intelligence · doi:10.1109/tpami.2016.2537325 · Cited by 37 · see citing works
“Similar to our approach it involves sampling image regions, but while they vector quantize the region descriptors, we propose a mechanism to select discriminative regions and build discriminative part based models from them.”
Multi-task learning for digital rock segmentation and characteristic parameters computation
Cao, D., Ji, S., Cui, R., et al. (2021) · Journal of Petroleum Science and Engineering · doi:10.1016/j.petrol.2021.109202 · Cited by 36 · see citing works
NOVA - A tool for eXplainable Cooperative Machine Learning
Heimerl, A., Baur, T., Lingenfelser, F., Wagner, J., & André, E. (2019) · International Conference on Affective Computing and Intelligent Interaction and workshops · doi:10.1109/acii.2019.8925519 · Cited by 36 · see citing works
“However, the performance of a ML-System is largely dependent on the amount and quality of the annotated training data [1], [2].”
BackgroundImagePairs: Realistic Super Resolution Dataset via Beam Splitter Camera Rig
Joze, H. R. V., Zharkov, I., Powell, K. D., Ringler, C., Liang, L., Roulston, A., Lutz, M., & Pradeep, V. (2020) · IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) · doi:10.1109/cvprw50498.2020.00267 · Cited by 33 · see citing works
“As proven, the input data itself plays a significant role in the machine learning processes [4, 78], especially deep learning approaches which are data hungry.”
BackgroundTraining Deformable Part Models with Decorrelated Features
Girshick, R., & Malik, J. (2013) · IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2013.375 · Cited by 32 · see citing works
“Surprisingly, no previous work has studied how DPM detection accuracy varies as a function of the number of negative training examples ( cf. [27]).”
BackgroundVisual Object Recognition with 3D-Aware Features in KITTI Urban Scenes
Yebes, J. J., Bergasa, L. M., & García-Garrido, M. Á. (2015) · Sensors · doi:10.3390/s150409228 · Cited by 33 · see citing works
“…of object orientation on 3D scenes [12], detection at far distances [13], determining the geometric layout of the scene [14], dealing with varying illumination conditions [15], appropriate modeling and parametric learning of complex scenes [16] and the generation of naturalistic datasets.”
BackgroundZelinsky, G. J., Peng, Y., Berg, A., & Samaras, D. (2013) · Journal of Vision · doi:10.1167/13.3.30 · Cited by 34 · see citing works
Kumar, A. A., Laer, L. V., Alaerts, M., Ardeshirdavani, A., Moreau, Y., Laukens, K., Loeys, B., & Vandeweyer, G. (2018) · Bioinformatics · doi:10.1093/bioinformatics/bty079 · Cited by 33 · see citing works
Object Detection Networks on Convolutional Feature Maps
Ren, S., He, K., Girshick, R., Zhang, X., & Sun, J. (2015) · arXiv · doi:10.48550/arxiv.1504.06066 · Cited by 454 · see citing works
“PM+CNN methods [24 ,11 29] and R-CNN/SPPnet. The DPM model is a less powerful classifier than a multi-layer network, whether applied to deep convolutional features or shallow HOG features. The work in [34] shows that a DPM classifier has diminishing improvement when given more training data. To method trained on mAP are…”
MethodDomain Adaptation for Visual Recognition
Gopalan, R., Li, R., Patel, V. M., & Chellappa, R. (2015) · Foundations and Trends® in Computer Graphics and Vision · doi:10.1561/0600000057 · Cited by 33 · see citing works
Yoga Pose Detection Using Posenet and k-NN
Shah, D., Rautela, V., Sharma, C., et al. (2021) · International Conference on Computing, Communication and Green Engineering (CCGE) · doi:10.1109/ccge50943.2021.9776451 · Cited by 32 · see citing works
Active learning for structured probabilistic models with histogram approximation
Sun, Q., Laddha, A., & Batra, D. (2015) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr.2015.7298984 · Cited by 30 · see citing works
Detection of eardrum abnormalities using ensemble deep learning approaches
Şenaras, Ç., Moberly, A. C., Teknos, T. N., Essig, G., Elmaraghy, C., Taj‐Schaal, N., Yua, L., & Gürcan, M. N. (2018) · SPIE Medical Imaging: Computer-Aided Diagnosis · doi:10.1117/12.2293297 · Cited by 31 · see citing works
Denkena, B., Bergmann, B., & Stoppel, D. (2020) · Journal of Manufacturing and Materials Processing · doi:10.3390/jmmp4030062 · Cited by 30 · see citing works
Suh-Lee, C., Jo, J.-Y., & Kim, Y. (2016) · IEEE Conference on Communications and Network Security · doi:10.1109/cns.2016.7860492 · Cited by 30 · see citing works
“The relationship between training data size and performance appears to be highly dependent on the types of datasets being studied [28, 29].”
BackgroundWhy do linear SVMs trained on HOG features perform so well?Preprint
Bristow, H., & Lucey, S. (2014) · arXiv (Cornell University) · doi:10.48550/arxiv.1406.2419 · Cited by 45 · see citing works
“[23] talk about the saturation of HOG at length, noting that more data sometimes decreases its performance.”
BackgroundSpeaker Anonymization for Personal Information Protection Using Voice Conversion Techniques
Yoo, I.-C., Lee, K., Leem, S.-G., Oh, H., Ko, B., & Yook, D. (2020) · IEEE Access · doi:10.1109/access.2020.3035416 · Cited by 29 · see citing works
“Deep learning-based algorithms, which are prominently used for speech recognition, require large quantities of training data [2], [3].”
BackgroundScaling Laws for Deep LearningPreprint
Rosenfeld, J. S. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2108.07686 · Cited by 26 · see citing works
Aodha, O. M., Stathopoulos, V., Brostow, G., Terry, M. A., Girolami, M., & Jones, K. E. (2014) · Proceedings - International Conference on Pattern Recognition/Proceedings/International Conference on Pattern Recognition · doi:10.1109/icpr.2014.12 · Cited by 25 · see citing works
“We know that not all data points contribute the same amount of information [62], [63].”
BackgroundSong, Y., Cai, W., Huang, H., Feng, D. D., Wang, Y., & Chen, M. (2016) · BMC Bioinformatics · doi:10.1186/s12859-016-1318-9 · Cited by 25 · see citing works
Factors of Transferability for a Generic ConvNet RepresentationPreprint
Azizpour, H., Razavian, A. S., Sullivan, J., Maki, A., & Carlsson, S. (2014) · arXiv (Cornell University) · doi:10.48550/arxiv.1406.5774 · Cited by 20 · see citing works
“Zhu et al.[48] suggest that increasing data is less effective than increasing the complexity of models or richness of representation and the former is prone to early performance saturation.”
BackgroundTraining Deformable Object Models for Human Detection Based on Alignment and Clustering
Drayer, B., & Brox, T. (2014) · Lecture notes in computer science · doi:10.1007/978-3-319-10602-1_27 · Cited by 23 · see citing works
“[22, 23], where landmark annotation, respectively human clustering, is used.”
BackgroundEfficient object detection using convolutional neural network-based hierarchical feature modeling
Lee, B., Erdenee, E., Jin, S., & Rhee, P. K. (2016) · Signal Image and Video Processing · doi:10.1007/s11760-016-0962-x · Cited by 21 · see citing works
Supervised learning and evaluation of KITTI's cars detector with DPM
Yebes, J. J., Bergasa, L. M., Arroyo, R., & Lazaro, A. (2014) · IEEE Intelligent Vehicles Symposium · doi:10.1109/ivs.2014.6856452 · Cited by 18 · see citing works
“These challenges may include and are not limited to object detection under occlusion [10], [11], estimation of objects orientation on 3D scenes [12], detection at far distances [13], determining geometric layout of the scene [14], [15], appropriate modeling and parametric learning of complex scenes [16] and large-enou…”
BackgroundVisual Object Detection Using Cascades of Binary and One-Class Classifiers
Çevıkalp, H., & Triggs, B. (2017) · International Journal of Computer Vision · doi:10.1007/s11263-016-0986-2 · Cited by 20 · see citing works
“…for object detection in the sense that they need a fine tuning of the error penalty parameter C (small changes of this parameter causes a large drop in accuracy), and increasing training set size may decrease the detection performance instead of an improvement in the accuracy (Zhu et al. 2012).”
BackgroundRefined particle swarm intelligence method for abrupt motion tracking
Lim, M. K., Chan, C. S., Monekosso, D., & Remagnino, P. (2014) · Information Sciences · doi:10.1016/j.ins.2014.01.003 · Cited by 19 · see citing works
“Motivated by the meta-level question prompted in [42] on whether there is a need to have more training data or better models for object detection, we raise similar question in this domain; will continued progress in visual tracking be driven by the increased complexity of tracking algorithms? As indicated in the earli…”
BackgroundEmpirical Upper Bound in Object Detection and MorePreprint
Borji, A., & Iranmanesh, S. M. (2019) · arXiv (Cornell University) · doi:10.48550/arxiv.1911.12451 · Cited by 18 · see citing works
“Some other related works in this line include [38, 67, 63].”
BackgroundDepth-based hand pose estimation: methods, data, and challengesPreprint
Supančič, J. S., Rogez, G., Yang, Y., Shotton, J., & Ramanan, D. (2015) · arXiv (Cornell University) · doi:10.48550/arxiv.1504.06378 · Cited by 19 · see citing works
“We specifically train an exemplar part model (EPM) constrained to model deformations consistent with 3D exemplars [43], which will be described further in a tech report.”
InfluentialSelf-citationDetecting Objects Using Deformation Dictionaries
Hariharan, B., Zitnick, C. L., & Dollár, P. (2014) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr.2014.256 · Cited by 17 · see citing works
“A larger D allows one to capture a larger number of poses, even ones not seen at train time [27].”
BackgroundParsing Based on Parselets: A Unified Deformable Mixture Model for Human Parsing
Dong, J., Chen, Q., Huang, Z., et al. (2015) · IEEE Transactions on Pattern Analysis and Machine Intelligence · doi:10.1109/tpami.2015.2420563 · Cited by 17 · see citing works
“Thus both appearance and structure parameters can be learned in a unified framework, which is critical for achieving the state-of-the-art performance for many applications [1], [2], [59].”
A Framework to Estimate the Key Point Within an Object Based on a Deep Learning Object Detection
Kurdthongmee, W., Suwannarat, K., & Wattanapanich, C. (2023) · HighTech and Innovation Journal · doi:10.28991/hij-2023-04-01-08 · Cited by 15 · see citing works
“Prior research has emphasized improving the scale, resolution, and diversity of the training data [3, 11-19].”
MethodData Summarization with Social Contexts
Zhuang, H., Rahman, R., Hu, X., et al. (2016) · doi:10.1145/2983323.2983736 · Cited by 14 · see citing works
“Sometimes additional training data could result in worse performance [37], and it is challenging to identify those data that are truly useful.”
BackgroundOptimizing 1-Nearest Prototype Classifiers
Wohlhart, P., Köstinger, M., Donoser, M., et al. (2013) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr.2013.66 · Cited by 14 · see citing works
“Also, the recent advances on the PASCAL visual object recognition challenge can largely be attributed to the steady increase in training database sizes [19].”
BackgroundInstrument Tracking with Rigid Part Mixtures Model
Węsierski, D., Wojdyga, G., & Jezierska, A. (2016) · Lecture notes in computer science · doi:10.1007/978-3-319-29965-5_3 · Cited by 14 · see citing works
“Consequently, our approach leverages successful flexible part mixtures model [6] that can be trained with datasets of modest size [17].”
MethodFusing subcategory probabilities for texture classification
Song, Y., Cai, W., Li, Q., Zhang, F., Feng, D. D., & Huang, H. (2015) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr.2015.7299070 · Cited by 13 · see citing works
“Fusion of subcategory results is typically performed by max or mean pooling [46, 15, 1, 47], and second layer of classifier learning [19, 12].”
BackgroundMethodActive learning using a self-correcting neural network (ALSCN)
Ilić, V., & Tadić, J. M. (2021) · Applied Intelligence · doi:10.1007/s10489-021-02515-y · Cited by 13 · see citing works
“Addition of the examples of classes that are already successfully recognized would not significantly increase the accuracy of the predictions [5] [6].”
BackgroundUnsupervised Total Variation Loss for Semi-supervised Deep Learning of Semantic SegmentationPreprint
Javanmardi, M., Sajjadi, M. S. M., Liu, T., et al. (2016) · arXiv (Cornell University) · doi:10.48550/arxiv.1605.01368 · Cited by 13 · see citing works
“[3] claim that improvements have a correlation not only with the methodologies used but also the amount of ar X iv :1 60 5.”
MethodEstimating Fruit Crop Yield through Deep Learning
Yu, H., Song, S., Ma, S., & Sinnott, R. (2019) · IEEE/ACM International Conference on Big Data Computing, Applications and Technologies (BDCAT) · doi:10.1145/3365109.3368766 · Cited by 13 · see citing works
“However, other round green shapes such as leaves may be misclassified as a green orange or a lime [9].”
Background
- Authors
- Zhu, X., Vondrick, C., Ramanan, D., & Fowlkes, C.
- Affiliations
- University of California, Irvine (US)
- Venue
- Procedings of the British Machine Vision Conference 2012
- Publisher
- British Machine Vision Association
- Published
- 2012
- Pages
- 80.1-80.11
- OpenAlex
- W2046875449
- Cited by
- 178
- References
- 0
- Citations by year
- 2026: 1 · 2025: 1 · 2024: 3 · 2023: 6 · 2022: 6 · 2021: 11 · 2020: 9 · 2019: 9 · 2018: 18 · 2017: 12 · 2016: 20 · 2015: 28 · 2014: 33 · 2013: 20 · 2012: 1
- FWCI
- 18.62
- Type
- conference paper
- Language
- en
- Open access
- closed
- Locations
- British Machine Vision Conference (BMVC) (publishedVersion)
- Topics
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Image and Video Retrieval Techniques
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Image and Object Detection Techniques
- Keywords
- object detection, training data size, model complexity, Bayes risk, parameter sharing, regularization, linear classifiers, feature space representation
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
