CoDeF - experiments

A smaller, simpler, and faster version of https://github.com/qiuyu96/CoDeF

I built this to understand capabilities and limitations of the approach

• batched training (much faster ~1min train)
• only 1x canonical/warping model (no background models)
• no masks
• optical flow computed with cv::calcOpticalFlowFarneback (rather than RAFT)
• no config files

Train

python3 run.py train --image_dir ./beauty_1

Train with a frame as the canonical image

python3 run.py train --image_dir ./beauty_1 --canonical ./beauty_1/00001.png

Generate frames

python3 run.py generate --checkpoint ./checkpoints/step=200.pt

Generate frames with a new canonical image

python3 run.py generate --checkpoint ./checkpoints/step=200.pt --canonical canonical.png