search

zhaoweicai / mscnn

Caffe implementation of our multi-scale object detection framework
405 stars 211 forks source link

Issues on the selection of training data #38

Closed du0002in closed 7 years ago

du0002in commented 7 years ago

Hi, Zhaowei

I noticed that when carrying out the training, you did not use the ground truth bounding boxes with occlusion greater and equal to 2. May I know the reason why ground truth bounding boxes with occlusion equal to 2 are excluded from the training. When occlusion equals to 2, the boxes are under the difficult category. If they are excluded from the training set, how to make sure that they can be revealed in the testing process? Did you see any drawbacks when taking those boxes into the training process?

Thanks.

lixincn2015 commented 7 years ago

I have the same question.

zhaoweicai commented 7 years ago

Hi @du0002in @lixincn2015 Sorry for late reply. It is kind of standard to exclude highly occluded objects from training set. The objects with occlusion>=2 are highly occluded, maybe only 10% are visible. Those instances are more background than foreground. Occlusion=1 (0%~50% occlusion) is somehow reasonable. I did try to use all instances, but the performance for Moderate is impaired and for Hard is increased. Anyway, you can choose your own dataset depending your needs.

du0002in commented 7 years ago

Thanks for the clarification.