Awesome

CANF-VC: Conditional Augmented Normalizing Flows for Video Compression

Update (09.11.23): CANF-VC++: Enhancing Conditional Augmented Normalizing Flows for Video Compression with Advanced Techniques

• We present CANF-VC++, an improved video compression framework from CANF-VC. CANF-VC++ demonstrates substantial Bjøntegaard-Delta rate savings of 40.2%, 38.1%, and 35.5% on UVG, HEVC Class B, and MCL-JCV datasets, respectively over CANF-VC. Please check our paper on arxiv for the details of our improvements. You can find the inference commands in full commands section.
• Performance

  • BD-rate (Anchor: CANF-VC ; GOP=32 ; only compress first 96 frames in each testing sequence)

    • 

  • R-D curves:

Update (08.30.22): CANF-VC with Error Propagation Aware Training Strategy

• CANF-VC-EPA is an enhanced version of CANF-VC. With exactly the same network architecture as CANF-VC, CANF-VC-EPA additionally introduces the Error Propagation Aware (EPA) training strategy from Guo et al., ECCV'20.
• Usage: Exactly the same as CANF-VC
• Performance

  • BD-rate (GOP=32 ; anchor: x265 veryslow). The best performer is marked in red and the second best in blue.

    • 

  • R-D curves:

  • <img src="https://user-images.githubusercontent.com/108980934/187449482-a9fad0fe-2506-47f2-8106-f4ada30d9ef7.png" width="400"> <img src="https://user-images.githubusercontent.com/108980934/187449448-aacf46ea-801f-48f7-8ac9-ce137ecc16ca.png" width="400"> <img src="https://user-images.githubusercontent.com/108980934/187574039-7f9b9c7c-2a45-4039-b0d2-10525c5c908e.png" width="400">

Project Installation

1. Prepare PyTorch 1.4.0 environment and correspond torchvision
2. Run sh install.sh
3. (Only needed for CANF-VC*) Install libbpg: https://github.com/mirrorer/libbpg 3.1 Configure path to libbpg as libbpg_path in dataloader.py
4. Download model weights & prepare testing data
5. Start evaluation: action=test/compress/decompress

Model Weight

• Get download link after filling out this form: https://forms.gle/Twsx8NofBjGakueL6

  Should be put under this project directory

Dataset

• Prepare all of your video sequence (in .png format), or
• Download all datasets:
  • Including:
    • U for UVG dataset
    • B, C, D, E for HEVC-B, -C, -D, -E dataset
    • M for MCL-JCV dataset
• We provide yuv2png.py for you to turn .yuv video into .png video frames
  • python yuv2png.py
  • Please specify the path & dataset to be converted in the file

Examples

• CANF-VC++ (PSNR):

  • test: $ python3 test.py --Iframe=ANFIC --Pframe=CANFVC_PP --MENet=PWC --motion_coder_conf=./CANF_VC/config/DVC_motion.yml --cond_motion_coder_conf=./CANF_VC/config/CANF_motion_quadtree_context.yml --residual_coder_conf=./CANF_VC/config/CANF_inter_coder_quadtree_context.yml --dataset=U --seq=Beauty --seq_len=96 --dataset_path=./video_dataset --lmda=2048 --model_dir=./models/CANF-VC++/PSNR --action=test --GOP=32
  • compress: $ python3 test.py --Iframe=ANFIC --Pframe=CANFVC_PP --MENet=PWC --motion_coder_conf=./CANF_VC/config/DVC_motion.yml --cond_motion_coder_conf=./CANF_VC/config/CANF_motion_quadtree_context.yml --residual_coder_conf=./CANF_VC/config/CANF_inter_coder_quadtree_context.yml --dataset=U --seq=Beauty --seq_len=96 --dataset_path=./video_dataset --lmda=2048 --model_dir=./models/CANF-VC++/PSNR --bitstream_dir=./bin/CANFVC_PP --action=compress --GOP=32
  • decompress: $ python3 test.py --Iframe=ANFIC --Pframe=CANFVC_PP --MENet=PWC --motion_coder_conf=./CANF_VC/config/DVC_motion.yml --cond_motion_coder_conf=./CANF_VC/config/CANF_motion_quadtree_context.yml --residual_coder_conf=./CANF_VC/config/CANF_inter_coder_quadtree_context.yml --dataset=U --seq=Beauty --seq_len=96 --dataset_path=./video_dataset --lmda=2048 --model_dir=./models/CANF-VC++/PSNR --bitstream_dir=./bin/CANFVC_PP --action=decompress --GOP=32

• CANF-VC (PSNR):

  • test: $ python3 test.py --Iframe=ANFIC --MENet=PWC --motion_coder_conf=./CANF_VC/config/DVC_motion.yml --cond_motion_coder_conf=./CANF_VC/config/CANF_motion_predprior.yml --residual_coder_conf=./CANF_VC/config/CANF_inter_coder.yml --dataset=D --dataset_path=./video_dataset --lmda=2048 --model_dir=./models/CANF-VC/PSNR --action=test --GOP=32
  • compress: $ python3 test.py --Iframe=ANFIC --MENet=PWC --motion_coder_conf=./CANF_VC/config/DVC_motion.yml --cond_motion_coder_conf=./CANF_VC/config/CANF_motion_predprior.yml --residual_coder_conf=./CANF_VC/config/CANF_inter_coder.yml --dataset=D --seq=BQSquare --seq_len=100 --dataset_path=./video_dataset --lmda=2048 --model_dir=./models/CANF-VC/PSNR --bitstream_dir=./bin --action=compress --GOP=32
  • decompress: $ python3 test.py --Iframe=ANFIC --MENet=PWC --motion_coder_conf=./CANF_VC/config/DVC_motion.yml --cond_motion_coder_conf=./CANF_VC/config/CANF_motion_predprior.yml --residual_coder_conf=./CANF_VC/config/CANF_inter_coder.yml --dataset=D --seq=BQSquare --seq_len=100 --dataset_path=./video_dataset --lmda=2048 --model_dir=./models/CANF-VC/PSNR --bitstream_dir=./bin --action=decompress --GOP=32

• CANF-VC* (PSNR):

  • test: $ python3 test.py --Iframe=ANFIC --MENet=PWC --motion_coder_conf=./CANF_VC/config/DVC_motion.yml --cond_motion_coder_conf=./CANF_VC/config/CANF_motion_predprior.yml --residual_coder_conf=./CANF_VC/config/CANF_inter_coder.yml --dataset=D --dataset_path=./video_dataset --lmda=2048 --model_dir=./models/CANF-VC_star/PSNR --action=test --GOP=32
  • compress: $ python3 test.py --Iframe=BPG --MENet=PWC --motion_coder_conf=./CANF_VC/config/DVC_motion.yml --cond_motion_coder_conf=./CANF_VC/config/CANF_motion_predprior.yml --residual_coder_conf=./CANF_VC/config/CANF_inter_coder.yml --dataset=D --seq=BQSquare --seq_len=100 --dataset_path=./video_dataset --lmda=2048 --model_dir=./models/CANF-VC_star/PSNR --bitstream_dir=./bin --action=compress --GOP=32
  • decompress: $ python3 test.py --Iframe=BPG --MENet=PWC --motion_coder_conf=./CANF_VC/config/DVC_motion.yml --cond_motion_coder_conf=./CANF_VC/config/CANF_motion_predprior.yml --residual_coder_conf=./CANF_VC/config/CANF_inter_coder.yml --dataset=D --seq=BQSquare --seq_len=100 --dataset_path=./video_dataset --lmda=2048 --model_dir=./models/CANF-VC_star/PSNR --bitstream_dir=./bin --action=decompress --GOP=32

Full Commands

• CANF-VC++:
  • test: $ python3 test.py --Iframe=ANFIC --Pframe=CANFVC_PP --MENet=PWC --motion_coder_conf=./CANF_VC/config/DVC_motion.yml --cond_motion_coder_conf=./CANF_VC/config/CANF_motion_quadtree_context.yml --residual_coder_conf=./CANF_VC/config/CANF_inter_coder_quadtree_context.yml --dataset={U/B/C/D/E/M} --dataset_path=/path/to/video_dataset --seq=SEQUENCE_TO_BE_COMPRESS(Optional) --seq_len=NUMBER_OF_FRAMES_TO_BE_COMPRESSED(Optional) --lmda={2048/1024/512/256} --model_dir=/path/to/CANF-VC++/{PSNR/MS-SSIM} --action=test --GOP=32
  • compress/decompress: $ python3 test.py --Iframe=ANFIC --Pframe=CANFVC_PP --MENet=PWC --motion_coder_conf=./CANF_VC/config/DVC_motion.yml --cond_motion_coder_conf=./CANF_VC/config/CANF_motion_quadtree_context.yml --residual_coder_conf=./CANF_VC/config/CANF_inter_coder_quadtree_context.yml --dataset={U/B/C/D/E/M} --seq=SEQUENCE_TO_BE_COMPRESS(Optional) --seq_len=NUMBER_OF_FRAMES_TO_BE_COMPRESSED(Optional) --dataset_path=/path/to/video_dataset --lmda={2048/1024/512/256} --model_dir=/path/to/CANF-VC++/{PSNR/MS-SSIM} --bitstream_dir=./bin --action={compress/decompress} --GOP=32
• CANF-VC:
  • test: $ python test.py --Iframe=ANFIC --MENet=PWC --motion_coder_conf=./CANF_VC/config/DVC_motion.yml --cond_motion_coder_conf=./CANF_VC/config/CANF_motion_predprior.yml --residual_coder_conf=./CANF_VC/config/CANF_inter_coder.yml --dataset={U/B/C/D/E/M} --dataset_path=/path/to/video_dataset --seq=SEQUENCE_TO_BE_COMPRESS(Optional) --seq_len=NUMBER_OF_FRAMES_TO_BE_COMPRESSED(Optional) --lmda={2048/1024/512/256} --model_dir=/path/to/CANF-VC/{PSNR/MS-SSIM} --action=test --GOP=32 {--msssim}
  • compress/decompress: $ python test.py --Iframe=ANFIC --MENet=PWC --motion_coder_conf=./CANF_VC/config/DVC_motion.yml --cond_motion_coder_conf=./CANF_VC/config/CANF_motion_predprior.yml --residual_coder_conf=./CANF_VC/config/CANF_inter_coder.yml --dataset={U/B/C/D/E/M} --seq=SEQUENCE_TO_BE_COMPRESS(Optional) --seq_len=NUMBER_OF_FRAMES_TO_BE_COMPRESSED(Optional) --dataset_path=/path/to/video_dataset --lmda={2048/1024/512/256} --model_dir=/path/to/CANF-VC/{PSNR/MS-SSIM} --bitstream_dir=./bin --action={compress/decompress} --GOP=32 {--msssim}
• CANF-VC*:
  • test: $ python3 test.py --Iframe=BPG --MENet=PWC --motion_coder_conf=./CANF_VC/config/DVC_motion.yml --cond_motion_coder_conf=./CANF_VC/config/CANF_motion_predprior.yml --residual_coder_conf=./CANF_VC/config/CANF_inter_coder.yml --dataset={U/B/C/D/E/M} --dataset_path=/path/to/video_dataset --seq=SEQUENCE_TO_BE_COMPRESS(Optional) --seq_len=NUMBER_OF_FRAMES_TO_BE_COMPRESSED(Optional) --lmda={2048/1024/512/256} --model_dir=/path/to/CANF-VC_star/{PSNR/MS-SSIM} --action=test --GOP=32 {--msssim}
  • compress/decompress: $ python3 test.py --Iframe=BPG --MENet=PWC --motion_coder_conf=./CANF_VC/config/DVC_motion.yml --cond_motion_coder_conf=./CANF_VC/config/CANF_motion_predprior.yml --residual_coder_conf=./CANF_VC/config/CANF_inter_coder.yml --dataset={U/B/C/D/E/M} --seq=SEQUENCE_TO_BE_COMPRESS(Optional) --seq_len=NUMBER_OF_FRAMES_TO_BE_COMPRESSED(Optional) --dataset_path=/path/to/video_dataset --lmda={2048/1024/512/256} --model_dir=/path/to/CANF-VC_star/{PSNR/MS-SSIM} --bitstream_dir=./bin --action={compress/decompress} --GOP=32 {--msssim}
• CANF-VC Lite:
  • test: $ python test.py --Iframe=ANFIC --MENet=SPy --motion_coder_conf=./CANF_VC/config/DVC_motion.yml --cond_motion_coder_conf=./CANF_VC/config/CANF_motion_predprior_Lite.yml --residual_coder_conf=./CANF_VC/config/CANF_inter_coder_Lite.yml --dataset={U/B/C/D/E/M} --dataset_path=/path/to/video_dataset --seq=SEQUENCE_TO_BE_COMPRESS(Optional) --seq_len=NUMBER_OF_FRAMES_TO_BE_COMPRESSED(Optional) --lmda={2048/1024/512/256} --model_dir=/path/to/CANF-VC-Lite/{PSNR/MS-SSIM} --action=test --GOP=32 {--msssim}
  • compress/decompress: $ python test.py --Iframe=ANFIC --MENet=SPy --motion_coder_conf=./CANF_VC/config/DVC_motion.yml --cond_motion_coder_conf=./CANF_VC/config/CANF_motion_predprior_Lite.yml --residual_coder_conf=./CANF_VC/config/CANF_inter_coder_Lite.yml --dataset={U/B/C/D/E/M} --seq=SEQUENCE_TO_BE_COMPRESS(Optional) --seq_len=NUMBER_OF_FRAMES_TO_BE_COMPRESSED(Optional) --dataset_path=/path/to/video_dataset --lmda={2048/1024/512/256} --model_dir=/path/to/CANF-VC-Lite/{PSNR/MS-SSIM} --bitstream_dir=./bin --action={compress/decompress} --GOP=32 {--msssim}

Citation

If you find this work useful for your research, please cite:

@article{canfvc,
  title={CANF-VC: Conditional Augmented Normalizing Flows for Video Compression},
  author={Ho, Yung-Han and Chang, Chih-Peng and Chen, Peng-Yu and Gnutti, Alessandro and Peng, Wen-Hsiao},
  journal={European Conference on Computer Vision},
  year={2022}
}