facebookresearch / multipathnet

A Torch implementation of the object detection network from "A MultiPath Network for Object Detection" (https://arxiv.org/abs/1604.02135)
Other
1.34k stars 275 forks source link

An error that may lead to a high precision drop. #44

Open samson-wang opened 7 years ago

samson-wang commented 7 years ago

Hi guys,

I have trained a model by multipathnet version frcnn on my own dataset with only one foreground class. Compared to the original frcnn model, which has AP at 0.93 on the same training and validation dataset, the multipath version only get AP around 0.7.

After many debug, I think the problem is that the coordinates preprocess mistakes. I change the corresponding code, then the AP go high to 0.94. I will give more details later.

samson-wang commented 7 years ago

https://github.com/samson-wang/multipathnet/commit/522a894fb899b4132db1b555cfdc6cd05976e693#diff-2ebe969b75ddaebef2a30717b15de628

This trick get the high AP result. At least for faster-rcnn model.