SparseR-CNN

SparseR-CNN copied to clipboard

train w/o pre-trained backbone

3

Hi, thank you for this excellent work first! Have you tried to train the model from scratch? I observed some strange things: 1. the mAP is much lower than reported...

StarsMyDestination

question about pred_bboxes.detach() in DynamicHead.forward()

2

In the forward method of DynamicHeadm, pred_bboxes are detached from graph before being used as input of next iteration, but proposal_features are kept. Why not just detach both or keep...

burningisnaive

Confused by the interaction between ROI features and proposal features

5

Thx for your interesting work and the code. I am confused by some points as follow when fusing ROI feats and proposal feats:  In this setting, the transformed proposal...

YueLiao

Question regarding proposal feature

4

I have a question regarding proposal feature. In DETR paper, reshaped feature map (HW x C) is given as input to transformer encoder to learn correlation between each pixels. However,...

abeyang00

Why the network needs proposal features/boxes?

3

Thanks for your great work! I have a few questions. As described in the paper, the proposal feature is used as a sparse representation for obtaining objects from the feature...

cfzd

issues about Mix-Precision training

1

Thanks for your great work!! Because of the limitation of GPU, I used pytorch native amp for training. However, I found some problems with numerical stability while computing giou, NaN...

hammer422

question about learned proposal boxes

1

Hi I have a question about the visualization shown in Figure 5. The learned proposal boxes in Figure 5 are quite different to the parameter (init_proposal_boxes.weight) in r101_100pro_3x_model.pth. For init_proposal_boxes.weight,...

liyunsheng13

how to detect mang pictures one time?

1

in demo.py?how to change the input from one image to a folder? demo = VisualizationDemo(cfg) if args.input: if len(args.input) == 1: args.input = glob.glob(os.path.expanduser(args.input[0])) assert args.input, "The input path(s) was...

henbucuoshanghai

how to change the confidence score in train net.py when eval-only?

1

I have tried to modify the SCORE_THRESH_TEST/ INSTANCES_CONFIDENCE_THRESH in defaults.py, but it doesn't work and the APs and other metrics don't change.

yeloiss

after trained about 7000 iters.it has error.

2

Traceback (most recent call last): File "projects/SparseRCNN/train_net.py", line 141, in args=(args,), File "/home/SparseR-CNN/detectron2/engine/launch.py", line 62, in launch main_func(*args) File "projects/SparseRCNN/train_net.py", line 129, in main return trainer.train() File "/homeSparseR-CNN/detectron2/engine/defaults.py", line 419,...

henbucuoshanghai