cyh-0 / cavp Goto Github PK

Hi, it seems that only the code for AVSBench-Semantic has been released. Could you please also release the code for AVSBench-Object? Thank you!

epoch 0: l_ce 0.396 l_ctr_av 4.032: 100%|███████████████████████████████████████| 4190/4190 [1:06:57<00:00, 1.04it/s]
0%| | 0/1437 [00:03<?, ?it/s]
Traceback (most recent call last):
File "/home/friends/yinhao/code/CAVP/main_vpo_stereo.py", line 282, in
main(0, hyp_param.gpus, hyp_param)
File "/home/friends/yinhao/code/CAVP/main_vpo_stereo.py", line 237, in main
trainer.validation(model_v, model_a, epoch, test_loader)
File "/home/friends/yinhao/anaconda3/envs/QwenAudio/lib/python3.9/site-packages/torch/utils/_contextlib.py", line 115, in decorate_context
return func(*args, **kwargs)
File "/home/friends/yinhao/code/CAVP/trainer/trainer_cavp_vpo_stereo.py", line 316, in validation
(
ValueError: too many values to unpack (expected 2)

https://github.com/cyh-0/CAVP/blob/main/utils/eval_utils.py#L33
I observed that your calculations of evaluation metrics (mIoU, F-score) during testing are fundamentally different from previous works, such as AVSBench, AVSegformer and so on. For example, in your code, the miou is calculated as (i1+i2+...+in)/(u1+u2+...+un), and the previous method is (i1/u1+i2/u2+...+in/un)/n. Trough our experiments, your calculation methods of evaluation metrics improves the results significantly compared with previous methods.
Above all, is it unfair in your paper to report SOTA using different calculations of evaluation metrics?

Hello, can you release the multi source dataset you created? Thank you!

Please refer to the Supplementary Material for details on the balancing process. where is it?