The results of test_net and demo have a huge different.

[BiwenW]

I have a custom dataset which only has only one category. And the dataset is pretty small (training around 150 images). When I tried Faster RCNN with 500 epoch. I got mAP= AP = 0.8806. However, after I run demo, it could only show couple of detection boxes. I realized that may be I need to adjust the number of boxes, so I modified ./lib/model/utils/net_utils.py line 52 to 150 followed by #876. It comes up couple more but way less than I want. I am wondering why I have this high mAP but not too much good detection bounding boxes. Also, how could I set the conf threshold? I checked the images after running demo, it seems like the probability is around above 70%. Any thought will be appreciated!

[Ivan-Tang-3D]

I have reviewed the demo and test.py. I found that the image would be preprocessed into a 'feature map',but in the test it would not be processed and directly passed into model.So demo had worse result.