pytorch-retinanet
pytorch-retinanet copied to clipboard
yhenon
Pytorch implementation of RetinaNet object detection.
Why use 0.05 for `score_threshold`. ?? Other detectors like `YOLO` uses 0.5. Can anybody explain?
I think sigmoid is not an appropriate choice to use in case of multiple classes in classification branch
Hello, how to assign GPU card number for training
I changed the backbone and train it, only get 2.9mAP with 10 epochs 's training. Can anyone share ResNet50 training history? Thanks!
RuntimeError: module must have its parameters and buffers on device cuda:0 (device_ids[0]) but found one of them on device: cpu I tried model = torch.load(model_path,map_location=torch.device('cpu')) and scores, classification, transformed_anchors =...
Does this loss implementation work for MultiClass classification? I see the use of Binary Cross Entropy, and it makes me wonder how it's being used. Is this applying the One...
In anchors.py file, line 110 and 11: > shift_x = (np.arange(0, shape[1]) + 0.5) * stride > shift_y = (np.arange(0, shape[0]) + 0.5) * stride Can somebody explain why we...
In the original paper, anchors are generated from feature maps on pyramid levels P3 to P7. Could someone enplane this? Or this is a bug?
Thank you for your implementation of retina-net detection! :) I found a little bit curious part about the "shift" function of anchors in my opinion anchor boxes' coordinate is center...
