Closed xiadingZ closed 5 years ago
I don't know what is your training param setting, but mine is showed on #2 . The features is too large to saved in .m with scipy.io for ShanghaiTech dataset, so I haven't use it to train on matlab. The number of objects I got on ShanghaiTech is about 190000.
I implement in pytorch and follow your settings and get 61.9 auc on shanghaitech, using only rgb. can you show your auc of it? If I use three-stream, training of svm will terminate because of excessive memory usage, on my server. I use pytorch's torchvision fasterrcnn_resnet50_fpn model to detect object, what's your number of objects on shanghaitech? Thanks very much