objectdetection-saliency-maps
objectdetection-saliency-maps copied to clipboard
RuoyuChen10
Is it possible for D-RISE to generate heatmaps for mask R-CNN?
Hi, I see you use Yolo v3 to do many tests. Is it possible to apply your D-RISE to the mask R-CNN?
Hi, I see you use Yolo v3 to do many tests. Is it possible to apply your D-RISE to the mask R-CNN?
Yes, you can have a try, load mask-cnn config, it should does work.
Okay. Thanks for your reply. You mean maybe directly use mask-cnn it can work? I will try it.
Okay. Thanks for your reply. You mean maybe directly use mask-cnn it can work? I will try it.
Yes, the only possible note is that Mask R-CNN will return masks, maybe a little note because methods like yolo v3 will only return boxes.
Hi, Is the D-RISE very slow? When I use 5000 nmask, use Mask R-CNN. It takes about 220s each object. Very slow
Hi, Is the D-RISE very slow? When I use 5000 nmask, use Mask R-CNN. It takes about 220s each object. Very slow
Yes, in the paper, D-RISE runs in approximately 70s per image (for all detections) for YOLOv3 and 170s for Faster R-CNN on NVidia Tesla V100.
Yes, I also see it. So slow. Very confused about how did they finish all MS COCO valid images if they use Faster R-CNN.....
Yes, I also see it. So slow. Very confused about how did they finish all MS COCO valid images if they use Faster R-CNN.....
When inference, each image can only choose one proposal for evaluation (Point Game).
You mean they only use 1 object for each image to evaluate (Point Game)?
One proposal has the highest prediction score.