Yuren Cong
Yuren Cong
Hi, Action Genome has no validation split. We use the same train/test split as the original AG (They provide "train/test" labels in the annotations)
Hi, someone emailed me a similar question, here is my reply: I wrote something for the dataset visualization when I got the demo from the author of Action Genome. Maybe...
If my memory is correct, these error rates are the accuracy. There are many true negative that could be considered as positive (with low confidence scores) making the accuracy low.
sorry for the late reply but it seems the image link is invalid. I would like to ask if these results are based on your reproduced version or our pre-trained...
I know what the problem might be. In fact, we've found that the larger batch size is not always better. We use bs=2 for the training and use multiple GPUs...
OpenImages V6 do have more object categories. However, for the VRD tasks, many categories do not appear in the relationship annotation. For data processing, please check BGNN. I remember that...
并不是所有600类都有relaitonship annotation的，举个例子，taxi在object detection task里有标注，但是并没有对应的triplet taxi-rel-obj或者sub-rel-taxi。因此就会把这种category滤掉
我记得是因为action genome有很多bounding box同时被标注成不同的class (比如book和notebook/food和sandwich等等)
Please check your CUDA version < 11.0
Hi, I am not so sure what you would like to do. If you can have such a result, you should know which frame the triplet belongs to.