Eric Liu
Eric Liu
You can reference this code , and modify the reorg layer source code https://github.com/eric612/MobileNet-YOLO/blob/master/src/caffe/layers/yolov3_layer.cpp#L287-L289
誤差可能由下列幾項造成的 * input image channel order (BGR or RGB ) * input image scale and mean * batch norm eps * relu and leakly relu * image scaler mode (keep...
Hi @jerryho-quanta , 我印象中darknet的image order 與caffe是相反的,你可以確認看看。
@jerryho-quanta , 如果你能從darknet2caffe轉換時,就把weighting調換當然可以,只是需要比較多的時間修改跟確認,或是你可以先改code將input image channel 調換試試 https://github.com/eric612/MobileNet-YOLO/blob/master/examples/yolo/yolo_detect.cpp#L474-#L480
@jerryho-quanta https://github.com/eric612/MobileNet-YOLO/blob/master/src/caffe/data_transformer.cpp 有兩種方式: 1. 在每個 DecodeDatumToCVMat 後面用opencv api 倒換順序 2. 直接修改 https://github.com/eric612/MobileNet-YOLO/blob/master/src/caffe/data_transformer.cpp#L792-L794 如果有segmentation 還要另外修改
如果只有2%的話 應該不只有order問題 太低了 你要確定看看這下列項目 https://github.com/eric612/MobileNet-YOLO/issues/177#issuecomment-541410096
[train.txt](https://github.com/eric612/MobileNet-YOLO/files/3976489/train.txt) [solver.txt](https://github.com/eric612/MobileNet-YOLO/files/3976490/solver.txt) [test.txt](https://github.com/eric612/MobileNet-YOLO/files/3976491/test.txt) [train_moblienetv2_yolov3_coco.sh.txt](https://github.com/eric612/MobileNet-YOLO/files/3976500/train_moblienetv2_yolov3_coco.sh.txt)
何謂垃圾數據?且你的LOG看起來應該是因為pre-trained weight不對
Please use imagenet pre-trained weights https://drive.google.com/file/d/0B3gersZ2cHIxZi13UWF0OXBsZzA/view Reference from https://github.com/chuanqi305/MobileNet-SSD
It looks like a resolution problem , the default mean , scale and input resolution were different between bdd100k and cityscapes model . The default parameter is 'bdd100k' in detect_and_segment.py