search

ruotianluo / pytorch-faster-rcnn

pytorch1.0 updated. Support cpu test and demo. (Use detectron2, it's a masterpiece)
MIT License
1.82k stars 475 forks source link

cannot get foreground and background rois in ./lib/layer_utils/proposal_target_layer.py #63

Open hedes1992 opened 6 years ago

hedes1992 commented 6 years ago

In the line 138-139 of ./lib/layer_utils/proposal_target_layer.py In my own dataset(car data from kitti-detection, just 2-class, car and background), at the beginning of this end2end training, it meet a small car, and all the overlaps between rois and this gt_bbox is lower than 0.01, so there are no foreground and background rois. current code just use a pdb, but no post-processing. I guess this batch might be discarded? Is there any other better method to avoid this? Thank you

But I ran the code on coco and pascal voc, there're no bug like this. I guess the reason is there are multiple classes...

ruotianluo commented 6 years ago

If there is a car in the image but no roi match that one, then you should consider changing the anchor size maybe.

hedes1992 commented 6 years ago

@ruotianluo Yes, I change the thresh to 0.0, and get a good result on training set. Thanks image

Saharkakavand commented 4 years ago

@hedes1992 I trained faster rcnn with kitti dataset, could you let me know how do you evaluate it?

hedes1992 commented 4 years ago

@hedes1992 I trained faster rcnn with kitti dataset, could you let me know how do you evaluate it?

I can only remember this repo can do test on kitti, then use the results to plot the PR curve by official toolkit from kittidevkir_object.zip, which is from the 2d-object detection webpage

Saharkakavand commented 4 years ago

@hedes1992 thank you.