why network seems to have a lot of false negatives ?

dineshreddy91 / Occlusion_Net

[CVPR2019]Occlusion-Net: 2D/3D Occluded Keypoint Localization Using Graph Networks

Other

139 stars 34 forks source link

why network seems to have a lot of false negatives ? #5

Closed jediofgever closed 4 years ago

dineshreddy91 commented 4 years ago

Can you explain with some examples??

moshanATucsd commented 4 years ago

Hi, thanks for sharing the code! we tried your method on KITTI, and it seems not able to handle occluded cars on the left, and the false positive (false car in middle) also exists

dineshreddy91 commented 4 years ago

This is a threshold problem in the detections from Maskrcnn(its currently 0.6 if you increase it the false positives in the middle will vanish). Since we saw only occlusions with a maximum of two overlaps in the carfusion dataset we cannot yet fix multiple overlap of cars as you show in your example on the left. Improving the detector will help such situations.

moshanATucsd commented 4 years ago

I see, thanks!