Pedestrian Detection

We plan to use Caltech Pedestrian Dataset with new annotations, CityPersons (a part of CityScapes) and KITTI for benchmarking.

Recent Update

2019.09.18 preview version of model v1 for Caltech Pedestrian Dataset is released.

Brief Introduction to Model Version

v1 - is designed for Caltech Pedestrian Dataset, covering pedestrian scale [30, 320]. It has 4 branches. Please check ./symbol_farm/symbol_structures.xlsx for details.

Inference Latency

Platform info: NVIDIA Jetson NANO, CUDA 10.0, CUDNN 7.5.0, TensorRT 5.1.6

Model Version	160×140	320×240	640×480	1280×720
v1	6.90ms(144.83FPS)	11.87ms(84.24FPS)	36.95ms(27.06FPS)	106.23ms(9.41FPS)
v2	-	-	-	-

Platform info: NVIDIA Jetson TX2, CUDA 10.0, CUDNN 7.5.0, TensorRT 5.1.6 (power mode: MAXN)

Model Version	160×140	320×240	640×480	1280×720	1920×1080
v1	3.63ms(275.43FPS)	6.80ms(147.36FPS)	15.87ms(63.01FPS)	43.33ms(23.08FPS)	93.93ms(10.65FPS)
v2	-	-	-	-	-

Platform info: NVIDIA RTX 2080TI, CUDA 10.0, CUDNN 7.4.2, TensorRT 5.1.5.0

Model Version	320×240	640×480	1280×720	1920×1080	3840×2160	7680×4320
v1	1.01ms(985.71FPS)	1.55ms(644.93FPS)	3.26ms(306.77FPS)	6.50ms(153.76FPS)	24.58ms(40.68FPS)	99.71ms(10.03FPS)
v2	-	-	-	-	-	-

Platform info: NVIDIA GTX 1060(laptop), CUDA 10.0, CUDNN 7.4.2, TensorRT 5.1.5.0

Model Version	320×240	640×480	1280×720	1920×1080	3840×2160
v1	1.25ms(800.00FPS)	2.93ms(341.80FPS)	7.46ms(134.08FPS)	16.03ms(62.39FPS)	62.80ms(15.92FPS)
v2	-	-	-	-	-

CAUTION: The latency may vary even in the same setting.

Accuracy on Caltech Pedestrian Dataset

After investigating the data, we found that Caltech Pedestrian Dataset is not well annotated, even giving the new annotations (not annotated, not aligned well, the highly occluded are annotated). The final data used for training: 1559 pos images (at least one pedestrian inside), 2691 neg images; 4786 pedestrian in total; the longer side of bboxes varies from 10 pixels to 500 pixels.

Download links for packed training and test sets:

Quantitative Results on Test Set

Currently, the quantitative results are not prepared well. We will release later.

Some Qualitative Results on Test Set

(we found that false positives are often appear in the small scales, probably due to noisy training instances. For large scales, v1 performs well.)

To play with the trained v1 model, please check ./accuracy_evaluation/predict.py.

User Instructions

Please refer to README in face_detection for details.