gdwang08

Hi, I did not receive any e-mail after filling in the survery form. Could you check it again? Thanks.

A very nice work! I am wondering if you have updated the results using the standard division of CIFAR100 super-class.

> Thanks for the fruitful discussion. > I got an accuracy of 94% on UCSDped2 using the pretrained model ckpt, the only difference I had with @Wolfybox 's dataloader is...

@hjs2027864933 You need a frame-by-frame extraction of the original video (.avi) for both training and testing and use our script (gen_patches.py) to generate spatio-temporal cubes (.npy format for each cube)....

@20220201yyds Correctly! You need to extract the video frames from the .avi videos from the downloaded STC before running gen_patches.py to generate cubes for training and testing. The extracted frames...

@hjs2027864933 The logs are correct for normal training. We have a function for loading .npy objects ([line](https://github.com/gdwang08/Jigsaw-VAD/blob/2708ab59145da8b8b53628c3b68a122188892a22/dataset.py#L150)). We access the image directory only for acquiring the frame length for each...

@20220201yyds If you strictly follow the gen_patches.py to generate cubes (sample_num = 9 for STC by default), you need to set sample_num=9 as well.

@hjs2027864933 The best performance is generally got in the middle of training ([60, 80] epochs in my case). According to my experiments, I get [83.5, 84.5] on average.

@summersnowfish First, you need to extract the frames from avi. files and structure the .jpg files as [this](https://github.com/gdwang08/Jigsaw-VAD/issues/2#issuecomment-1279702893). Then you could run gen_patches.py to generate .npy files.

@summersnowfish "shanghaitech_train_detect_result_yolov3.pkl" stores the frame-by-frame object detection results for STC by YOLOv3. You may first check if the number of extracted frames match the number of frame-level detection results for...