Neural Residual Flow Fields for Efficient Video Representations

1. Download MPI sintel dataset

Download MPI sintel dataset from here

2. GMA optical flow estimator

To obtain optical flow estimations for pretraining, we are using GMA from here. Note that it dose not have to do with our identity.

3. Training

Training neural residual flow fields (NRFF)

# frame 0 - 6
python train_video_flow_midkey.py --use-estimator --lr 0.0005 --training-step 30000 --data-dir {sintel dataset training directory} --video-name alley_1 --start-frame 0 --num-frames 7 --jpeg-quality 98 --hidden-features 96 --use-estimator --tag start0_jq98_hf96
# frame 7 - 13
python train_video_flow_midkey.py --use-estimator --lr 0.0005 --training-step 30000 --data-dir {sintel dataset training directory} --video-name alley_1 --start-frame 7 --num-frames 7 --jpeg-quality 98 --hidden-features 96 --use-estimator --tag start7_jq98_hf96
# frame 14 - 20
python train_video_flow_midkey.py --use-estimator --lr 0.0005 --training-step 30000 --data-dir {sintel dataset training directory} --video-name alley_1 --start-frame 14 --num-frames 7 --jpeg-quality 98 --hidden-features 96 --use-estimator --tag start14_jq98_hf96
# frame 21 - 27
python train_video_flow_midkey.py --use-estimator --lr 0.0005 --training-step 30000 --data-dir {sintel dataset training directory} --video-name alley_1 --start-frame 21 --num-frames 7 --jpeg-quality 98 --hidden-features 96 --use-estimator --tag start21_jq98_hf96

Training baseline (SIREN)

python train_video.py --data-dir {sintel dataset training directory} --video-name alley_1 --hidden-features 256 --num-frames 28 --lr 0.001 --training-step 30000 --tag baseline_siren_hf256