Active Learning for Deep Detection Neural Networks

“To this end, we follow [3] and resize the logits spatially using the bilinear upsampling.”

“er must have the same size as the image. To this end, we follow [3] and resize the logits spatially using the bilinear upsampling. Each Fire Residual module follows the same architecture indicated in [12]. Also, each downsampling layer is composed of convolution and pooling layers which are applied in parallel and con…”

“[14] and Gal et al .”

“CaltechPed[11] • 640x480 resolution • 51363 Images • 20062 Boxes • Videos”

“In other words, there is a domain shift [19, 25, 28, 27] between Xl and Xu that makes the active learning procedure more challenging.”

“In other words, there is a domain shift [19, 25, 28, 27] between Xl and Xu that makes the active learning procedure more challenging.”

“[2] computed the marginal score [23] of candidate bounding boxes and integrate them using different merging functions.”

“Also, each downsampling layer is composed of convolution and pooling layers which are applied in parallel and concatenated at the output [17].”

“[1] proposed a new image-level scoring process to rank unlabeled images for automatic selection.”

“Another, more practical, solution is to exploit temporal coherence in image sequences to enforce sampling diversity [43].”

“The general goal of AL is to leverage a scoring mechanism (uncertainty [13, 46], entropy [1, 36], heuristics [24, 27], set-selection [57], clustering [5]) which can identify samples most likely to be useful for future training.”

“For example, AL-OD [11] calculates pixel-level scores and uses the scores to select the informative samples.”

“The main objective of these works was to either perform active learning mainly to reduce the labelling and training cost (Gal, Islam, and Ghahramani 2017; Choi et al. 2021; Roy, Unmesh, and Namboodiri 2018; Aghdam et al. 2019; Haussmann et al. 2020; Yu et al. 2021; Feng et al. 2019), or to improve the training mechani…”

“To improve the heuristic function, recent work (Lakshminarayanan et al., 2016) (Alling-ham et al., 2021) on explicit ensembles shows strong results for uncertainty computing, and (Aghdam et al., 2019) show that adding temporal reasoning can be beneficial for data selection on object detection task.”

“Recent works can also be found on active learning for DNN based object detectors [54], [55], where the focus is on adaptations of active learning to existing object detection frameworks.”

“the training of CNNs have been proposed [3], [4].”

“A salient example of active learning being used in a detection task is in (Aghdam et al., 2019) where the neural network’s mispredictions of bounding boxes in an object detection task is used to query human labelers for frame-level annotations.”

“It is widely applied across different fields such as image recognition (Budd et al., 2021; Feng et al., 2019; Aghdam et al., 2019), natural language processing (Priya et al., 2019; Shardlow et al., 2019; Maldonado & Harabagiu, 2019), and drug discovery (Chen & Gu, 2020; Reker et al., 2017;…”

“40,41 42–46 Our previous work on LOCalization-Based Active Learning (LOCAL) draws on the most successful of these approaches.”

“[17] ve [18]’de, ESA’ların e ˘ gitimi için aktif ö ˘ grenme ile etiketlenmi¸s örneklerin sayısının azaltılmasını içeren yöntemler önerilmi¸stir.”

“b) Active Learning for Object Detection: In the context of active learning for object detection, specific metrics related to characteristics of the underlying network can be applied [28].”

“Other uncertainty-based methods attempt to aggregate pixel-level scores for each image [57], reformulate detectors by adding Bayesian inference to estimate the uncertainty [58] or replace conventional detection head with the Gaussian mixture model to compute aleatoric and epistemic uncertainty [13].”

“[2] proposed an Image-level score that accumulated pixel-level importance scores such that an image with the highest number of false-positive and false negative predictions would be sent to the oracle.”

“Earliest attempts in active learning for object detection include LL4AL [1] adapting instance loss predictions and AL4DeepDetection [13] combining uncertainty metrics for foreground objects and background pixels.”

“The key component in active learning is the selection strategy, with uncertainty being the most widely used due to its broad applicability [33], [34].”

“At present, active learning has been embedded into many machine learning algorithms [15]– [18] and applied to many fields, such as image recognition [2], [19]–[21], text classification [22], [23], visual question answering [24] and object detection [25]–[27].”

“It has been widely used across various domains [27–29].”

“In 2D visual tasks, Aghdam et al. [9] applied active learning by aggregating pixel-level scores.”

“We notice that different methods vary in sample selection strategy [6, 20, 2], i.e., which sample to be annotated, while few works target at designing efficient query method, which we consider is important for active learning in object detection.”

“The majority of existing AL methods [1, 2, 12, 31] rely on a combination of the criteria such as model uncertainty and the diversity in the labeled pool.”

“criterion based on this straightforward way, including uncertainty criteria: entropy [3], distance from the decision boundary [8], 1-vs-2 [9], uncertainty [2], and confidence [10], and committee criteria: disagreement between multiple copies of the model [3], disagreement between nearby boxes [10] and pixel neighborho…”

“Active learning [2], [3], [52], [54], which requires continuous human input to modify intermediate object detection results, incorporates human responses that are used directly as additional information.”

“Among those approaches limited to AL for DN include the work presented by Ducoffe and Precioso that uses margin theory to compute instances along the decision boundary, [4], [1], [8] and [9].”

“A typical example is to start with a model trained on a small subset of the data set and iteratively add more samples based on the models confidence on the remaining unlabeled samples [15].”

“Recent works can also be found on active learning for DNN based object detectors (Choi et al., 2021; Aghdam et al., 2019), where the focus is on adaptations of active learning to existing object detection frameworks.”

“A. Active Learning for 2D/3D Object Detection Active learning (AL) methods have been widely used to reduce annotation costs in both 2D object detection [19]– [25] and 3D object detection [1]–[3], [26]–[30].”

“As highlighted by [4], even using various automatic labeling tools still requires at least an average of 35 seconds to label a single pedestrian in an urban road scene.”

“Classical approaches rely on uncertainty-based sampling 7,8,11 (e.g., entropy or model confidence).”

“With regard to their applications, AL methods can be divided into two categories, i.e., task-aware AL [25], [26], [27], [28] and task-agnostic AL [29], [30].”

“Active learning has been extensively studied and applied to tasks such as image classification [89]–[91], 2D/3D object detection [50], [64], [72], [99], [100], and 2D semantic segmentation [101], [102].”

“Recent AL studies have applied AL to various deep learning applications such as image classification [22, 24, 6], object detection [1].”

“Recent advances in active learning for object detection include LL4AL [1], which adapts instance loss predictions, and Aghdam et al. [26], which combines uncertainty metrics for foreground objects and background pixels.”

“[21], [26] calculates pixel scores and use them for selecting informative samples.”

“- Deep active learning : Deep Active learning has been widely utilized in various fields, including image recognition [26], [27], [28] and object detection [29], [30], [31].”

“Aghdam et al [3] proposed a novel approach combining pixel-level scores to obtain an image-level score for doing active learning for the task of pedestrian detection task.”

“Another, more practical, solution is to exploit temporal coherence in image sequences to enforce sampling diversity [43].”

“Active learning is used to ask experts for providing annotation for unlabeled samples [43], [44].”

“На сегодняшний день существует немало статей [5-9], посвященных исследованию подходов активного обучения в задачах обнаружения объектов и сегментации изображений, которые показали свою эффективность.”

“Lately, WSL has been successfully applied to the object detection task [15–18], however their adoption in robotics is still at a preliminary stage.”