Skip to content

Latest commit

 

History

History
193 lines (153 loc) · 7.16 KB

README.md

File metadata and controls

193 lines (153 loc) · 7.16 KB

Dataset Downloading & Preprocessing

Assume the project to be located in (change as required)

REPO_ROOT=/path/to/monodepth_benchmark

Paths

Datasets are expected to be in $REPO_ROOT/data by default. Path management is done in src/paths.py. The expected datasets are

datas: dict[str, str] = {
    'kitti_raw': 'kitti_raw_sync',
    'kitti_raw_lmdb': 'kitti_raw_sync_lmdb',
    'kitti_depth': 'kitti_depth_benchmark',
    'syns_patches': 'syns_patches',
}

where keys are the dataset identifiers in the registry and values are the stems of the path where the dataset is stored. E.g. kitti_raw would be stored in $REPO_ROOT/data/kitti_raw_sync.

If you wish to store datasets in other locations, create the file $REPO_ROOT/PATHS.yaml. This file should not be tracked by Git, since it might contain sensitive information about your machine/server. Populate the file with the following contents:

# -----------------------------------------------------------------------------
MODEL_ROOTS: []

DATA_ROOTS:
  - /path/to/dataroot1
  - /path/to/dataroot2
  - /path/to/dataroot3
# -----------------------------------------------------------------------------

NOTE: Multiple roots may be useful if training in an HPC cluster where data has to be copied locally.

Replace the paths with the desired location(s). Once again, the dataset would be expected to be located in /path/to/dataroot1/kitti_raw_sync. These roots should be listed in preference order, and the first existing dataset directory will be selected.

NOTE: This procedure can be applied to change the default locations in which to find pretrained models. Simply add the desired paths to MODEL_ROOTS instead.

Download

Before downloading the datasets, create a placeholder directory for each benchmark dataset. This will prevent errors from the devkits thinking the datasets are unavailable.

cd $REPO_ROOT
mkdir -p data/{kitti_raw_sync,kitti_raw_sync_lmdb,kitti_depth_benchmark,syns_patches}

Kitti Raw Sync

This is the base Kitti Raw Sync dataset, including the training images and velodyne LiDAR. The zipfiles will be extracted and deleted automatically.

cd $REPO_ROOT/data
mkdir kitti_raw_sync; cd kitti_raw_sync
../../api/data/download_kitti_raw_sync.sh  # Ensure file is executable with `chdmod +x ...`

The expected dataset size, given by du -shc ./* is

62G     2011_09_26
24G     2011_09_28
6.0G    2011_09_29
51G     2011_09_30
43G     2011_10_03
184G    total

Kitti Depth Benchmark

This is the updated Kitti Benchmark dataset, including the corrected ground-truth depth maps. The zipfiles will be extracted and deleted automatically.

cd $REPO_ROOT/data
mkdir kitti_depth_benchmark; cd kitti_depth_benchmark
../../api/data/download_kitti_bench.sh  # Ensure file is executable with `chdmod +x ...`

python ../../api/data/export_kitti_depth_benchmark.py  # Copy benchmark data to `kitti_raw_sync`

The expected dataset size, given by du -shc ./* is

1.9G     depth_selection
160K     devkit
13G      train
1.1G     val
16G      total

The dowloaded version of kitti_depth_benchmark can be deleted after copying it to kitti_raw_sync. Otherwise, you might be able to use it to submit to the official Kitti Benchmark.

SYNS-Patches

The SYNS-Patches dataset can currently be downloaded from the MDEC CodaLab website.

NOTE: The ground-truth for both val and test sets are not publicly released! This is to prevent overfitting from repeated evaluation. If you wish to evaluate on this dataset, consider participating in the challenge!

cd $REPO_ROOT/data
../api/data/download_syns_patches.sh  # Ensure file is executable with `chdmod +x ...`

The expected dataset size, given by du -shc . (subdirs omitted for brevity) is

974M      total

Evaluation Targets

In this section we will generate the ground-truth depth targets used for evaluation. First, copy the splits provided in $ROOT/api/data to the corresponding dataset. Then export the targets for each test/val split and each dataset.

cd $ROOT
cp -r api/data/splits_kitti data/kitti_raw_sync/splits
cp -r api/data/splits_syns data/syns_patches/splits
api/data/export_targets.sh  # Ensure file is executable with `chdmod +x ...`

NOTE: The test/val set for SYNS-Patches are held out. To generate the predictions for your submission please refer to this section

Expected number of images:

  • Kitti Eigen (test): 697
  • Kitti Eigen Zhou (test): 700
  • Kitti Eigen Benchmark (test): 652
  • SYNS Patches (val): 400

Depth Hints (Optional)

If you wish to train using proxy depth supervision from SGBM predictions, generate them using the following commands. We follow the procedure from DepthHints and compute the hints using multiple hyperparameters and the min reconstruction loss.

api/data/export_hints.sh  # If you wish to process all splits below...

python api/data/export_kitti_hints.py --split eigen_zhou --mode train
python api/data/export_kitti_hints.py --split eigen_zhou --mode val

python api/data/export_kitti_hints.py --split eigen --mode train
python api/data/export_kitti_hints.py --split eigen --mode val

python api/data/export_kitti_hints.py --split eigen_benchmark --mode train
python api/data/export_kitti_hints.py --split eigen_benchmark --mode val

NOTE: This takes quite a while (overnight), so only generate hints for the splits you are going actively use. This will cause errors when running the tests, but you can check that the desired splits have all the expected hints.

NOTE: The script will check if a given hint already exists and skip it. So don't worry if you need to interrupt it and carry on with your life :)

Generate LMDBs (Optional)

Once you have finished downloading and preprocessing the Kitti dataset, you can optionally convert it into LMDB format. This should make the data load faster, as well as reduce the load on the filesystem. You might find this beneficial if you are training in a cluster with limited bandwidth.

cd $REPO_ROOT/data
mkdir kitti_raw_sync_lmdb; cd kitti_raw_sync_lmdb
python ../../api/data/export_kitti_lmdb.py  --use-hints 1 --use-benchmark 1   # Change if needed to skip hints.

NOTE: This takes quite a while (a few hours), but the script will check if a given database already exists and skip it. So don't worry if you need to interrupt it and carry on with your life :)

Tests

You can test that all datasets have been loaded/preprocessed correctly using:

cd $REPO_ROOT
python -m pytest tests/test_data/test_kitti_raw.py
python -m pytest tests/test_data/test_kitti_raw_lmdb.py
python -m pytest tests/test_data/test_syns_patches.py

NOTE: Some tests will fail unless all depth hints and benchmark depths have been processed. Ignore these errors if not needed.