README.md

Video Depth Estimation

See configs in configs/evaluation/videodepth.yaml.

python videodepth/infer.py
# torchrun --nnodes=1 --nproc_per_node=8 videodepth/infer_mp.py  # accelerate with multi gpus
python videodepth/eval.py              # align=scale&shift by default
python videodepth/eval.py align=scale  # override with align=scale

Infer to Generate .npy & .png Depth Files

Assume eval_models only have pi3, infer.py will generate folders like:

recons-eval
├── ...
├── outputs
|   └── videodepth
|       ├── hydra (runtime configs)
|       └── pi3
|           ├── bonn
|           |   ├── sequence_1
|           |   |   ├── _time.json
|           |   |   ├── xxxxxx.npy
|           |   |   ├── xxxxxx.png
|           |   |   └── ...
|           |   ├── sequence_2
|           |   └── ...
|           ├── kitti
|           ├── nyu-v2
|           └── sintel
└── ...

Eval with Generated Depth Files

After infer.py finishes, you can run eval.py to evaluate the results.

Then the videodepth metrics will be generated in outputs/videodepth/{dataset_name}-metric-{align}.csv. For example, if align=scale, then

recons-eval
├── ...
├── outputs
|   └── videodepth
|       ├── hydra (runtime configs)
|       ├── pi3
|       |   ├── bonn
|       |   |   ├── sequence_1
|       |   |   |   ├── _time.json
|       |   |   |   ├── xxxxxx.npy
|       |   |   |   ├── xxxxxx.png
|       |   |   |   └── ...
|       |   |   ├── sequence_2
|       |   |   └── ...
|       |   ├── kitti
|       |   └── sintel
|       ├── bonn-metric-scale.csv
|       ├── kitti-metric-scale.csv
|       └── sintel-metric-scale.csv
└── ...