search

dineshreddy91 / Occlusion_Net

[CVPR2019]Occlusion-Net: 2D/3D Occluded Keypoint Localization Using Graph Networks
Other
139 stars 34 forks source link

How to handle cars without annotations in training #14

Closed motokimura closed 4 years ago

motokimura commented 4 years ago

Hi, thank you for the code!

According to the following issue, some cars are not annotated in CarFusion dataset. https://github.com/dineshreddy91/carfusion_to_coco/issues/1

When training Occlusion_Net with CarFusion dataset, how did you handle them? It would be great if you could tell me where this handling is implemented in this repository.

Thank you in advance.

dineshreddy91 commented 4 years ago

Currently occlusion-net is a top down approach so you can take any good vehicle detector like maskrcnn or Hrnet and then use the current data from car fusion to train the key point head. We used the same methodology for the occlusion-net. But if you are using bottom up approaches like open pose. you need to mask the unannotated cars.

motokimura commented 4 years ago

Thank you for your quick reply.

So you mean you loaded pretrained maskrcnn weight at the beginning, and the detector part was frozen during training? I guess un-annotated vehicles will degrade detector performance without freezing.

dineshreddy91 commented 4 years ago

yes. if you do not freeze the weights of the detector the accuracy will be lower.

motokimura commented 4 years ago

Thank you! I got it now.