Infinite loop on the minimum detection per image

jwyang / graph-rcnn.pytorch

[ECCV 2018] Official code for "Graph R-CNN for Scene Graph Generation"

732 stars 157 forks source link

Infinite loop on the minimum detection per image #104

Closed khieu closed 3 years ago

khieu commented 3 years ago

I'm trying to run graph rcnn on my task with COCO image and I got an infite loop in lib/scene_parser/rcnn/modeling/roi_heads/box_head/inference.py with the output:

Number of proposals 1 is too small, retrying filter_results with score tresh = 0.0.

When checking the default config value, the minimum detection in each image is _C.MODEL.ROI_HEADS.MIN_DETECTIONS_PER_IMG = 2. I would really appreciate if anyone knows how to work around this issue. Thank you.

khieu commented 3 years ago

The cause for this issue is that I have the imdb_1024.h5 file created incorrectly. All images in the visual genomes dataset has to be extracted to the images folder, not under VGG_100k folder.