microsoft / DCVC

Deep Contextual Video Compression
MIT License
408 stars 66 forks source link

Severe quality degradation issue when using the DCVC-DC Intra model !!! #29

Open dotori71 opened 1 year ago

dotori71 commented 1 year ago

First of all, I want to express my gratitude for releasing both your code and pre-trained model weights for all the settings. The testing results closely align with the data reported in your paper, particularly regarding the UVG, HEVC-B, and MCL-JCV datasets.

Recently, I conducted tests on the intra model IntraNoAR of DCVC-DC using the provided pre-trained weights (cvpr2023_video_psnr.pth.tar) from this link. My tests involved using images from a different dataset in the RGB color space for research purposes. Unfortunately, during this process, I encountered issues that led to instability in my model during training.

It took some time for me to debug my implementation, and I finally found that the main issue stemmed from the degradation of the I-frame when compressed by your intra model. I didn't update any learnable parameters in the model which means the result should be the same as the pre-trained model you provided.

Here is the result of the degraded compressed image

All the images above were downloaded from the original video sources of the commonly used training dataset Vimeo90K, accessible via this link.

Could you please test your intra model with the provided images to check for any problems? Additionally, it would be greatly appreciated if you could provide an additional training script for both your intra-frame and inter-frame models.