EuroCity Persons: A Novel Benchmark for Person Detection in Traffic Scenes

The recall for pedestrian boxes of the reasonable scenario larger than 40 px is about 100% for all methods for an IoU of 0.5.

“The recall for pedestrian boxes of the reasonable scenario larger than 40 px is about 100% for all methods for an IoU of 0.5.”

Faster R-CNN all performs best overall.

“Faster R-CNN all , that is trained with pedestrians of all sizes and of occlusions up to 80%, performs best overall.”

Both one stage detectors YOLOv3 and SSD perform better than R-FCN at least on the reasonable and occluded test scenarios.

“Interestingly, both one stage detectors YOLOv3 and SSD perform better than R-FCN at least on the reasonable and occluded test scenarios.”

EuroCity Persons provides orientation information for all persons larger than 40 px, including those riding.

“We provide this information for all persons larger than 40 px (including those riding).”

• AssumptionRests on an assumption: Overall object orientation is an important cue for predicting future motion of persons in traffic scenes.

EuroCity Persons contains over 211200 person orientation annotations.

“The dataset furthermore contains a large number of person orientation annotations (over 211200).”

A substantial fraction of the person annotations in EuroCity Persons are unique.

“This means that a substantial fraction of the person annotations in the dataset are unique, although especially at traffic lights and in slow moving traffic, same persons might appear in different annotations.”

There is a considerable performance difference for small sized pedestrians, as expected by the lower box recall.

“As could be expected by the lower box recall shown in Figure 4, there is a considerable performance difference for small sized pedestrians.”

By driving through a large part of Europe, during all four seasons, in most weather conditions (apart from heavy rain or snowfall), and during day and night, EuroCity Persons recorded very diverse backgrounds and person appearances.

“By driving through a large part of Europe, during all four seasons, in most weather conditions (apart from heavy rain or snowfall), and during day and night, we recorded very diverse backgrounds and person appearances, see Table 5.”

• AssumptionRests on an assumption: Datasets captured in few countries and only in daylight and dry weather conditions hamper generalization to real-world applications.

Very crowded scenarios were collected in Lyon with on average 9.5 pedestrians per image.

“Very crowded scenarios have been collected in Lyon with on average 9.5 pedestrians per image.”

EuroCity Persons contains night-time scenes.

“It also contains night-time scenes.”

The EuroCity Persons dataset provides a large number of highly diverse, accurate and detailed annotations of pedestrians, cyclists and other riders in urban traffic scenes.

“In this paper, we introduce the EuroCity Persons dataset, which provides a large number of highly diverse, accurate and detailed annotations of pedestrians, cyclists and other riders in urban traffic scenes.”

• Also stated in: § 2.4 Main Contributions · p. 4 · plain

  “We introduce the EuroCity Persons dataset, which provides a large number of highly diverse, accurate and detailed annotations of persons (pedestrians, cyclists, and other riders) in urban traffic scenes across Europe.”

EuroCity Persons is nearly one order of magnitude larger than person datasets used previously for benchmarking, in terms of manual annotations.

“EuroCity Persons is nearly one order of magnitude larger than person datasets used previously for benchmarking.”

• Also stated in: § 1 INTRODUCTION · p. 2 · plain

  “EuroCity Persons is nearly one order of magnitude larger than person datasets used previously for benchmarking, in terms of manual annotations.”

Faster R-CNN variants trained and tested with original image resolution perform slightly worse for the reasonable and occluded test scenarios than other Faster R-CNN variants.

“The Faster R-CNN variants ( all origsize, baseline ) that are trained and tested with the original image resolution perform slightly worse for the reasonable and occluded test scenario than the other Faster R-CNN variants.”

EuroCity Persons contains over 238200 person instances manually labeled in over 47300 images.

“With over 238200 person instances manually labeled in over 47300 images, EuroCity Persons is nearly one order of magnitude larger than person datasets used previously for benchmarking.”

• Also stated in: § 1 INTRODUCTION · p. 2 · plain

  “With over 238200 person instances manually labeled in over 47300 images, EuroCity Persons is nearly one order of magnitude larger than person datasets used previously for benchmarking, in terms of manual annotations.”

Faster R-CNN all performs slightly better than Faster R-CNN occluded on the occluded test scenario.

“It also performs slightly better than Faster R-CNN occluded on the occluded test scenario.”

Faster R-CNN all evaluated on 1.3× upscaled images has higher recall for smaller pedestrians compared to Faster R-CNN all origsize.

“Boxes of Faster R-CNN all are evaluated on 1 . 3 × upscaled images in contrast to Faster RCNN all origsize resulting in a higher recall in particular for smaller pedestrians.”

Variants of the two stage method Faster R-CNN perform overall best on the three test scenarios.

“Variants of the two stage method Faster R-CNN perform overall best on the three test scenarios.”

The EuroCity Persons images were collected on-board a moving vehicle in 31 cities of 12 European countries.

“The images for this dataset were collected on-board a moving vehicle in 31 cities of 12 European countries.”

• Also stated in: § 1 INTRODUCTION · p. 2 · plain

  “The images for this dataset were collected on-board a moving vehicle in 31 cities of 12 European countries, see Figure 1 and Table 5.”

• Also stated in: § 3.1 Dataset Collection · p. 4 · plain

  “We collected the images of the EuroCity Persons dataset from a moving vehicle in 31 cities of 12 European countries.”

Faster R-CNN detection performance suffers from deleting and disturbing bounding box locations.

“The detection performance of Faster R-CNN suffers from deleting and disturbing the bounding box locations.”

Deleting 25% of bounding boxes results in a miss-rate of 11.7 for Faster R-CNN.

“Deleting 25% of the bounding boxes results in a miss-rate of 11.7.”

Processing rates for RFCN, Faster R-CNN, SSD and YOLOv3 on non-upscaled test images were 1.2 fps, 1.7 fps, 2.4 fps and 3.8 fps, respectively.

“Processing rates for the RFCN, Faster R-CNN, SSD and YOLOv3 on non-upscaled test images were 1.2 fps , 1.7 fps , 2.4 fps and 3.8 fps , respectively”

Pre-training with ECP is especially valuable for the hard or occluded cases, involving improvements of about 10 percentage points in LAMR.

“Pre-training with ECP is especially valuable for the hard or occluded cases, involving improvements of about 10 percentage points in LAMR .”

These benefits are larger than pre-training with CP.

“These benefits are larger than pre-training with CP.”

Detection performance deteriorates when not using ignore regions during training.

“In accordance with earlier findings [4], we observe that detection performance deteriorates when not using ignore regions during training.”

• cites [4] [4] S. Zhang, R. Benenson, and B. Schiele, 'CityPersons: A Diverse Dataset for Pedestrian Detection,' Proc. IEEE CVPR , 2017. - sourceNot checked yet

  Not checked in the free run.

Small and occluded pedestrians are a further common source for false negatives.

“Small and occluded pedestrians are a further common source for false negatives as already shown by the small and occluded test scenarios.”

The orientation annotation deviation lies within the requirements of Section 3.2.

“We note that this lies within the requirements of Section 3.2 as well.”

In qualitative analysis there are false negatives where the head is occluded.

“In our qualitative analysis we have false negatives where the head is occluded.”

An IoU threshold of 0.5 is not sufficient to suppress detections that have very diverse aspects.

“We use an IoU threshold of 0.5 which is not sufficient to suppress detections that have very diverse aspects.”

Table 11 shows that detection results decrease slightly for the multitask network with the Biternion loss.

“In contrast to [9] which shows that training multiple tasks together can improve the overall result, Table 11 shows that the detection results decrease slightly for the multitask network with the Biternion loss.”

• cites [9] [9] R. B. Girshick, 'Fast R-CNN,' in Proc. IEEE ICCV , 2015, pp. 14401448. - sourceNot checked yet

  Not checked in the free run.

The confusion of pedestrians and background is even stronger when OHEM is applied, as seen comparing R-FCN results with and without OHEM.

“This effect is even stronger when OHEM is applied as seen when comparing R-FCN results with and without OHEM.”

The logarithmic relation between training set size and detection performance also holds on our benchmark for Faster R-CNN and SSD.

“We observe that logarithmic relation between training set size and detection performance also holds on our benchmark for Faster R-CNN and SSD.”

The best results on CP were achieved by the Faster R-CNN model pretrained on ECP with a LAMR of 14.9.

“Again, the best results were achieved by the Faster R-CNN model pretrained on ECP with a LAMR of 14 . 9 .”

The average deviation of the bounding box extents stays below 1.4 px for objects up to 200 px high.

“Figure 10 shows that the average deviation of the bounding box extents stays below 1.4 px for objects up to 200 px high”

By adding riders as an additional class, pedestrian detection performance improves for the protocol which requires pedestrians to be classified as such (10.3 vs. 11.0).

“By adding riders as an additional class, one observes that the pedestrian detection performance improves for the protocol which requires pedestrians to be classified as such (10.3 vs. 11.0).”

Because of the high IoU between pedestrians, not all of them can be detected because of the greedy NMS.

“Because of the high IoU between pedestrians not all of them can be detected because of the greedy NMS.”

Pre-training results in a decrease of 2, 9, and 7 percentage points in LAMR for the reasonable, small, occluded CP validation datasets, respectively, when compared to using the original CP training data set.

“Similarly, pre-training results in a decrease of 2, 9, and 7 percentage points in LAMR for the reasonable, small, occluded CP validation datasets, respectively, when compared to using the original CP training data set.”

False negatives already occur for an IoU threshold of 0.5.

“These already occur for an IoU threshold of 0.5 as shown in the NMS repressing category.”

Faster RCNN achieved an average precision of 65.9 for pedestrians on the official KITTI test set for the moderate scenario.

“Faster RCNN achieved an average precision of 65.9 for pedestrians on the official KITTI test set for the moderate scenario.”

Our model pre-trained on ECP achieved an average precision of 72.6 on the KITTI test set.

“Our model pre-trained on ECP achieved an average precision of 72.6 on the KITTI test set.”

At a fppi of 0.3 for Faster R-CNN all on the reasonable setup, 230 missed annotations larger than 32 px were manually counted.

“At a fppi of 0.3 for Faster R-CNN all on the reasonable setup we manually count 230 missed annotation larger than 32 px .”

The model pretrained on ECP outperforms the CP based model with an average precision of 80.8.

“The model pretrained on ECP outperforms the CP based model with an average precision of 80 . 8 .”

There is only a slight difference in performance when the network is trained to regress a bounding box for the rider alone or for the rider including the ride type.

“There is only a slight difference in performance when the network is trained to regress a bounding box for the rider alone or for the rider including the ride type.”

Depictions and reflections are among the hardest error sources to take care off: at decreasing fppi the share of this error type increases.

“In terms of classification errors, depictions and reflections are among the hardest error sources to take care off: at decreasing fppi the share of this error type increases.”

Pre-training on ECP and fine-tuning on a smaller target dataset (KITTI, CP) yields better results than training solely on the target dataset.

“pre-training on ECP and fine-tuning (post-training) on a smaller target dataset (KITTI, CP) yields better results than training solely on the target dataset.”

CityPersons and EuroCity Persons differ regarding annotation bias, as the aspect ratios of all bounding boxes provided by CP are the same in contrast to ECP.

“Still they differ regarding annotation bias, as the aspect ratios of all bounding boxes provided by CP are the same in contrast to ECP (cf. Section 3.2).”

Training on daytime and testing on night-time gives significantly worse results than training and testing on the same time-of-day.

“Table 10 shows that training on daytime and testing on night-time gives significantly worse results than training and testing on the same time-of-day.”

Due to dataset biases, solely training on a dataset from the other domain without fine-tuning results in worse detection performance.

“However, due to the dataset biases solely training on a dataset from the other domain without fine-tuning results in worse detection performance.”

Even at training data sizes about one order of magnitude larger than existing ones, considered state-of-the-art deep learning methods (Faster R-CNN and SSD) do not saturate in detection performance.

“Even at the new training data sizes that are about one order of magnitude larger than previous ones, the considered deep learning methods do not saturate in detection performance.”

With 75% of training samples a LAMR of 10.0 is achieved.

“Note that with 75% of the training samples a LAMR of 10.0 is achieved (see Figure 8).”

Placing hallucinations at random locations only slightly influences overall detection performance.

“Placing hallucinations at random locations only slightly influences the overall detection performance.”

Of the touched groundtruth samples, an increasing proportion is non-matched for decreasing fppi.

“Of these touched groundtruth samples, an increasing proportion is non-matched, for decreasing fppi .”

Pre-training results in an increase of 5, 8, and 11 percentage points in average precision for the easy, moderate and hard KITTI validation datasets, respectively, when compared to using the original KITTI training data set.

“Pre-training results in an increase of 5, 8, and 11 percentage points in average precision for the easy, moderate and hard KITTI validation datasets, respectively, when compared to using the original KITTI training data set.”

For the reasonable and small test scenario the LAMR drops by about two points when not using ignore regions.

“For the reasonable and small test scenario the LAMR drops by about two points.”

Certain pedestrian poses and aspect ratios can lead to multiple detections for the same pedestrian.

“Certain pedestrian poses and aspect ratios can lead to multiple detections for the same pedestrian as shown in the Multidetections category.”

When using CP to pretrain for KITTI and vice versa, detection performance is worse than when using ECP for pre-training.

“On the other hand, when using CP to pretrain for KITTI and vice versa, the detection performance is worse than when using ECP for pre-training.”

Clothes, depictions and reflections are main sources for confusion with real pedestrians and thus for false positives.

“As can be seen, clothes, depictions and reflections are main sources for confusion with real pedestrians and thus for false positives”

These findings not only hold on the validation datasets.

“These findings not only hold on the validation datasets.”

Adding 10% hallucinations that more resemble pedestrians ('false positives') results in a more significant drop in performance of 3.3 points.

“Adding 10% hallucinations that more resemble pedestrians ('false positives') result in a more significant drop in performance of 3.3 points.”

Using the more conservative 30% figure, there are additional 99 missed annotations for pedestrians larger than 32 px, bringing the total missed annotation to 329.

“Using the more conservative 30% figure, we estimate that, in fact, there are additional 99 missed annotations for pedestrians larger than 32 px, bringing the total missed annotation to 329.”

Faster R-CNN pretrained on CP achieves an average precision of 77.5 on the moderate setting of the KITTI benchmark.

“For the moderate setting of the KITTI benchmark the Faster R-CNN pretrained on CP achieves an average precision of 77 . 5 .”

We included in this benchmark the orientation estimation of the overall body.

“In preparation for this, we included in this benchmark the orientation estimation of the overall body”

At equal training set sizes, night-time performance is worse (a LAMR of four points higher) than day-time.

“Comparing day- and night-time detection performance, one observes from Table 10 that at equal training set sizes, night-time performance is worse (a LAMR of four points higher).”

Night-performance is a few percentage points lower than day-time performance.

“The experiments also showed that night-performance is a few percentage points lower than day-time performance.”

In our case correct ignore region handling boosts performance by 1.5 points.

“In our case it boosts performance by 1 . 5 points (see Table 11).”

Experiments suggest there is a bias derived from large geographical region.

“Foremost, experiments suggest that there is indeed a bias derived from large geographical region.”

A visual inspection shows that remaining errors are indeed 'hard', even for a human.

“A visual inspection shows that the remaining errors are indeed 'hard', even for a human, see Table 6.”

The existence of a bias (geographical) is statistically significant with a confidence interval of 99%.

“We found that the existence of a bias is statistically significant with a confidence interval of 99%.”

These confusions result in high scoring false positives also for sizes larger than 80 px.

“These confusions result in high scoring false positives also for sizes larger than 80 px .”

Approximately 75% of false positives at 0.3 fppi analyzed in Figure 6 are smaller than 80 pixels.

“Approximately 75% of the false positives at 0.3 fppi analyzed in Figure 6 are smaller than 80 pixels.”

Depictions, reflections and clothes are often confused with real pedestrians.

“We show quantitatively in Figure 6 and qualitatively in Table 6 that depictions, reflections and clothes are often confused with real pedestrians.”

We have created the new EuroCity Persons dataset, which takes annotations of persons in urban traffic scenes to a new level in terms of quantity, diversity and detail.

“We have created the new EuroCity Persons dataset, which takes annotations of persons in urban traffic scenes to a new level in terms of quantity, diversity and detail.”

Performance still improves with increasing training set size with a decent gradient.

“Our study shows that performance still improves with increasing training set size with a decent gradient (i.e. Figure 8).”

The orientation estimation of overall body can be jointly trained with detection task at minimal performance loss.

“and showed that the latter can be jointly trained with the detection task at minimal performance loss.”

Detection rate increases to 97% at a distance of 25 m (object height 100 px).

“See Figure 11, the detection rate increases to 97% at a distance of 25 m (object height 100 px ).”

Correct ignore region handling has an impact on detection performance.

“As in [4] we show that a correct ignore region handling has an impact on detection performance.”

• cites [4] [4] S. Zhang, R. Benenson, and B. Schiele, 'CityPersons: A Diverse Dataset for Pedestrian Detection,' Proc. IEEE CVPR , 2017. - sourceNot checked yet

  Not checked in the free run.

Detectors pre-trained with new dataset and fine-tuned on target dataset yield superior performance than those trained on target dataset only (improvements on KITTI and CityPersons by 5-11 and 2-9 points, respectively).

“In experiments on transfer learning, we showed that detectors pre-trained with the new dataset and finetuned on a target dataset, yield superior performance than those trained on the target dataset only (improvements on KITTI and CityPersons by 5-11 and 2-9 points, respectively).”

Detectors improve their performance with decreasing distance (increasing object size).

“The detectors improve their performance with decreasing distance (increasing object size).”

Small objects are still very challenging despite great amount of small sized pedestrians in training dataset.

“In Figure 11 and in our baseline experiments we show that small objects are still very challenging despite the great amount of small sized pedestrians present in our training dataset.”

This difference is enlarged when the entire day- and night-time training sets of ECP are used.

“This difference is enlarged when the entire dayand night-time training sets of ECP are used as the former is an order of magnitude larger.”

When training detection network jointly for pedestrians and riders, confusions between two person classes can be reduced.

“When training a detection network jointly for pedestrians and riders we have already shown that confusions between the two person classes can be reduced.”

This is a larger difference than that between the performances using 75% and 100% of the training data in Figure 8.

“This is a larger difference than that between the performances using 75% and 100% of the training data in Figure 8.”

Jointly trained detection models for riders and pedestrians achieve a lower miss rate for the pedestrian class than models trained for pedestrians-only, when the precise class is enforced.

“In our experiments, the jointly trained detection models for riders and pedestrians achieve a lower miss rate for the pedestrian class, than models trained for pedestrians-only, when the precise class is enforced.”

Adding orientation estimation to object detection lowers detection performance by a single percentage point for the Biternion loss.

“Adding orientation estimation to object detection lowers the detection performance by a single percentage point for the Biternion loss.”

The EuroCity Persons detection benchmark was sampled by extracting every 80th frame without further filtering to limit selection bias.

“To limit selection bias [15], we extracted every 80-th frame for our detection benchmark without further filtering.”

• cites [15] [15] A. Torralba and A. A. Efros, 'Unbiased look at dataset bias,' in Proc. IEEE CVPR , June 2011, pp. 1521-1528. - sourceNot checked yet

  Not checked in the free run.

The validation set was placed intermittently with the training and test sets to avoid the latter two containing the same physical person.

“By placing furthermore the validation set intermittently with the training and test, it was all but avoided that the latter two would contain the same physical person.”

EuroCity Persons annotations are overall accurate due to the implemented quality control procedures.

“Finally, thanks to the implemented quality control procedures, annotations are overall accurate.”

• Also stated in: § 3.2 Dataset Annotation · p. 5 · plain

  “Annotations were double checked by a quality validation team that was disjoint from the annotation team. If needed, several feedback iterations were run between the teams to achieve a consolidated outcome.”

In terms of annotation quantity and data diversity, CityPersons and Tsinghua-Daimler Cyclist had, so far, the most to offer for the pedestrian and riders classes.

“In terms of the annotation quantity and data diversity, CityPersons [4] and TsinghuaDaimler Cyclist [5] had, so far, the most to offer for the pedestrian and the riders class.”

• cites [4] [4] S. Zhang, R. Benenson, and B. Schiele, 'CityPersons: A Diverse Dataset for Pedestrian Detection,' Proc. IEEE CVPR , 2017. - sourceNot checked yet

  Not checked in the free run.

• cites [5] [5] X. Li, F. Flohr, Y. Yang, H. Xiong, M. Braun, S. Pan, K. Li, and D. M. Gavrila, 'A New Benchmark for Vison-Based Cyclist Detection,' in IEEE Intell. Veh , 2016. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Sequences were split only at points where the recording vehicle had a speed larger than 7 km/h to make the test, validation, and training sets disjunct in terms of people.

“To facilitate that the test, validation and training sets are disjunct in terms of people we only splitted sequences at points in time where the recording vehicle had a speed larger than 7 km/h .”

Faster R-CNN can reach top performance for pedestrian detection if it is properly optimized.

“As shown in [4], it can reach top performance for pedestrian detection if it is properly optimized.”

• cites [4] [4] S. Zhang, R. Benenson, and B. Schiele, 'CityPersons: A Diverse Dataset for Pedestrian Detection,' Proc. IEEE CVPR , 2017. - sourceNot checked yet

  Not checked in the free run.

The generalization capabilities needed for transfer learning can be addressed to four dataset aspects: diversity, quantity, accuracy, and detail.

“We argue that the generalization capabilities needed for transfer learning shown in the last section can be addressed to four dataset aspects, namely, diversity, quantity, accuracy, and detail.”

Current deep learning methods need to adapt to these biases first, even if the same vision task is involved.

“Current deep learning methods need to adapt to these biases first, even if the same vision task is involved.”

Because of inherent dataset biases, models trained on one dataset have to be fine-tuned to be applicable on other datasets.

“Because of these inherent dataset biases models trained on one dataset have to be fine-tuned to be applicable on other datasets.”

Despite dataset biases, models pre-trained on ECP learned general features for pedestrian detection useful for other datasets after finetuning.

“Despite the dataset biases, the models were able to learn general features for the task of pedestrians detection when pre-trained on ECP which proves useful for other datasets as well after finetuning.”

With a confidence interval of 99%, the null-hypothesis (the non-existence of a regional bias) can be rejected for both regions.

“Hence, with a confidence interval of 99% , the null-hypothesis (the non-existence of a regional bias) can be rejected for both regions.”

The fact that saturation does not occur can be attributed to the diversity of the data.

“The fact that saturation does not occur can be attributed to the diversity of the data.”

Overall results are worse than those of other experiments due to the comparatively small training sets used.

“Overall results are worse than those of other experiments due to the comparatively small training sets used.”

There is a logarithmic relation between the amount of training data and the performance of deep learning methods.

“[7] shows a logarithmic relation between the amount of training data and the performance of deep learning methods.”

• cites [7] [7] C. Sun, A. Shrivastava, S. Singh, and A. Gupta, 'Revisiting unreasonable effectiveness of data in deep learning era,' Proc. IEEE ICCV , 2017. -Not checked yet

  Not checked in the free run.

Data still remains a driving factor for person detection performance in urban traffic scenes.

“A main outcome from the EuroCity Persons (ECP) experiments is that data still remains a driving factor for the person detection performance in urban traffic scenes”

• Also stated in: § 6 CONCLUSIONS · p. 16 · plain

  “The experiments show that data is still a driving factor for the person detection performance in urban traffic scenes”

Resampling of a dataset can be applied to evaluate the significance of benchmark results.

“[76] shows that resampling of a dataset can be applied to evaluate the significance of benchmark results.”

• cites [76] [76] J. Demˇ sar, 'Statistical comparisons of classifiers over multiple data sets,' Journal of Machine Learning Research , pp. 1-30, 2006. -Not checked yet

  Not checked in the free run.

Existing biases in current databases are too strong to allow avoiding a fine-tuning step on the target dataset altogether; this holds even for a much larger source training dataset like ECP.

“existing biases in current databases are too strong to allow avoiding a fine-tuning step on the target dataset altogether; this holds even for a much larger source training dataset like ECP.”

System performance regarding person detection in urban traffic settings has improved by an order of magnitude over the past decade.

“System performance regarding person detection in urban traffic settings has improved by an order of magnitude over the past decade; it is now closing in on human performance.”

Some false positives can be eliminated when taking advantage of known scene geometry constraints.

“Second, some false positives can be eliminated, when taking advantage of known scene geometry constraints (e.g. pedestrians or riders should be on the ground plane, their heights should be physically plausible when accounting for perspective mapping).”

In the evaluation protocol of [4], this case is not considered as riders are always handled as ignore regions.

“In the evaluation protocol of [4] this case is not considered as riders are always handled as ignore regions.”

• cites [4] [4] S. Zhang, R. Benenson, and B. Schiele, 'CityPersons: A Diverse Dataset for Pedestrian Detection,' Proc. IEEE CVPR , 2017. - sourceNot checked yet

  Not checked in the free run.

More person training data is especially helpful in non-standard poses, in rainy or nighttime conditions, or under partial occlusions.

“More person training data is especially helpful in non-standard poses, in rainy or nighttime conditions, or under partial occlusions.”

The way annotations are performed proves to be important.

“The way annotations are performed proves to be important as well.”

Current single-frame pedestrian detection performance lags that of an attentive human by an order of magnitude.

“A recent paper [1] finds that current single-frame pedestrian detection performance lags that of an attentive human by an order of magnitude.”

• cites [1] [1] S. Zhang, R. Benenson, M. Omran, J. Hosang, and B. Schiele, 'Towards reaching human performance in pedestrian detection,' IEEE TPAMI , vol. 34, no. 4, pp. 973-985, 2018. -Not checked yet

  Not checked in the free run.

Tracking can suppress such false positives ([11] shows a reduction by up to 37%).

“Tracking can suppress such false positives ( [11] shows a reduction by up to 37%).”

• cites [11] [11] M. Enzweiler and D. M. Gavrila, 'Monocular pedestrian detection: Survey and experiments,' IEEE TPAMI , vol. 31, no. 12, pp. 21792195, 2009. -Not checked yet

  Not checked in the free run.

There is a potential for substantial further performance improvement.

“Thus there is a potential for a substantial further performance improvement; an improvement which would be important with the advent of fully self-driving vehicles.”

A further doubling of current training size (110000 pedestrians) is projected to yield a reduction of LAMR from 9.4 to about 7.6 points.

“A further doubling of the current training size (110000 pedestrians) is projected to yield a reduction of the LAMR from 9.4 to about 7.6 points.”

It is beneficial to train specific detectors for classes that otherwise might be confused with the target class.

“We go beyond [4] to show that it is beneficial to train specific detectors for classes that otherwise might be confused with the target class.”

• cites [4] [4] S. Zhang, R. Benenson, and B. Schiele, 'CityPersons: A Diverse Dataset for Pedestrian Detection,' Proc. IEEE CVPR , 2017. - sourceNot checked yet

  Not checked in the free run.

Most top performing methods of common generic object detection benchmarks depend on a simple greedy NMS [77].

“Interestingly, most top performing methods of the common generic object detection benchmarks depend on a simple greedy non-maximum suppression (NMS) [77].”

• cites [77] [77] N. Bodla, B. Singh, R. Chellappa, and L. S. Davis, 'Soft-NMS improving object detection with one line of code,' in Proc. IEEE ICCV , 2017. - sourceNot checked yet

  Not checked in the free run.

It is now closing in on human performance.

“System performance regarding person detection in urban traffic settings has improved by an order of magnitude over the past decade; it is now closing in on human performance.”

More data remains part of the solution on how to improve performance.

“More data remains part of the solution on how to improve performance.”

This NMS poses a problem for overlapping objects e.g. in pedestrian groups.

“This NMS poses a problem for overlapping objects e.g. in pedestrian groups.”

The claim that state-of-the-art detection performance is far away from practical use for vehicle application is incorrect.

“State-of-the-art detection performance (e.g. correct detection around 90% at 0 . 1 -0 . 3 fppi ) is sometimes cited as evidence that performance is far away from practical use for the vehicle application. This is incorrect, as can be readily inferred from the fact that there are already several vision-based person detection systems on-board production vehicles on the market.”

• AssumptionRests on an assumption: There are already several vision-based person detection systems on-board production vehicles on the market.
• AssumptionRests on an assumption: Current commercial systems are in the context of driver assistance, meaning that a correct detection performance of about 90% is acceptable, as long as the false alarm rate is essentially zero.