Weak segmentation supervised deep neural networks for pedestrian detection

Highlights：

• Jointly optimizing detection and segmentation tasks helps to improve the detector.

• Weak segmentation masks automatically generated by depth maps are used for training.

• Fusion of classification and segmentation results further improves detection.

• We report state-of-the-art pedestrian detection performance on three RGBD datasets.

• We obtain segmentation results of good quality without using accurate annotations.