Download the full KITTI 3D Object detection dataset:
- left color images of object data set (12 GB)
- camera calibration matrices of object data set (16 MB)
- training labels of object data set (5 MB)
Download the nuScenes and Waymo datasets. Note- Both these datasets are quite big and requires about 1TB storage space each.
Setup KITTI, nuScenes and Waymo as follows:
./code
├── data
│ ├── KITTI
│ │ ├── ImageSets
│ │ ├── kitti_split1
│ │ ├── training
│ │ │ ├── calib
│ │ │ ├── image_2
│ │ │ └── label_2
│ │ │
│ │ └── testing
│ │ ├── calib
│ │ └── image_2
│ │
│ ├── nusc_kitti
│ │ ├── ImageSets
│ │ └── nuscenes
│ │ ├── maps
│ │ ├── samples
│ │ └── v1.0-trainval
│ │
│ └── waymo
│ ├── ImageSets
│ └── raw_data
│ ├── training
│ └── validation
│
├── experiments
├── images
├── lib
├── nuscenes-devkit
│ ...Simply put the soft links:
cd data/KITTI/
ln -sfn your_path/kitti/training training
ln -sfn your_path/kitti/testing testing
cd ../..Next download the patched nuScenes devkit. This patched devkit provides lidar points per box in KITTI format and also supports evaluation on nuScenes front camera:
git clone https://github.com/abhi1kumar/nuscenes-devkitThen follow the instructions at convert_nuscenes_to_kitti_format_and_evaluate.sh to get nusc_kitti_org folder. Finally link this folder with the following command:
cd data/nusc_kitti/
ln -sfn your_path/nusc_kitti_org nusc_kitti_orgThis generates the following structure:
./code
├── data
│ ├── nusc_kitti
│ │ ├── ImageSets
│ │ ├── nusc_kitti_org
│ │ │ ├── train
│ │ │ │ ├── calib
│ │ │ │ ├── image_2
│ │ │ │ └── label_2
│ │ │ │
│ │ │ └── val
│ │ │ ├── calib
│ │ │ ├── image_2
│ │ │ └── label_2
│ │ └── nuscenes
│ │ ├── maps
│ │ ├── samples
│ │ └── v1.0-trainval
│ │ Finally run the script setup_split.py to generate training and validation folders of the nuScenes Val split:
ln -sfn your_path/nusc_kitti_training_mapped training
ln -sfn your_path/nusc_kitti_validation_mapped validation
python setup_split.py
cd ../..The script uses soft-links for efficient storage and creates the desired directory structure for nuScenes Val, which can be used by any KITTI based detector.
./code
├── data
│ ├── nusc_kitti
│ │ ├── ImageSets
│ │ ├── nusc_kitti_org
│ │ │ ├── train
│ │ │ │ ├── calib
│ │ │ │ ├── image_2
│ │ │ │ └── label_2
│ │ │ │
│ │ │ └── val
│ │ │ ├── calib
│ │ │ ├── image_2
│ │ │ └── label_2
│ │ ├── nuscenes
│ │ │ ├── maps
│ │ │ ├── samples
│ │ │ └── v1.0-trainval
│ │ │
│ │ ├── training
│ │ │ ├── calib
│ │ │ ├── image
│ │ │ └── label
│ │ │
│ │ └── validation
│ │ ├── calib
│ │ ├── image
│ │ └── labelDecompress the Waymo zip files into their corresponding directories:
ls *.tar | xargs -i tar xvf {} -C your_target_dirEach directory contains tfrecords. Arrange them in the following fashion:
./code
├── data
│ └── waymo
│ ├── ImageSets
│ └── raw_data
│ ├── training
│ │ ├── segment-10017090168044687777_6380_000_6400_000_with_camera_labels.tfrecord
│ │ └── segment-10023947602400723454_1120_000_1140_000_with_camera_labels.tfrecord
│ │
│ └── validation
│ ├── segment-10203656353524179475_7625_000_7645_000_with_camera_labels.tfrecord
│ └── segment-1024360143612057520_3580_000_3600_000_with_camera_labels.tfrecordThen, setup the Waymo devkit. The Waymo devkit is setup in a different environment to avoid package conflicts with our DEVIANT environment:
# Set up environment
conda create -n py36_waymo_tf python=3.7
conda activate py36_waymo_tf
conda install cudatoolkit=11.3 -c pytorch
# Newer versions of tf are not in conda. tf>=2.4.0 is compatible with conda.
pip install tensorflow-gpu==2.4
conda install pandas
pip3 install waymo-open-dataset-tf-2-4-0 --userNext convert the segments to the KITTI format using converter.py. Note that we have commented out the code for saving lidar frames which takes huge amount of time. Make sure to keep the number of processes num_proc to the highest number supported by your GPU. The conversion takes about 3 days to complete on our end.
conda activate py36_waymo_tf
cd data/waymo/
python converter.py --load_dir "" --save_dir your_path/datasets/waymo_open_organized/ --split training --num_proc 10
python converter.py --load_dir "" --save_dir your_path/datasets/waymo_open_organized/ --split validation --num_proc 10
ln -sfn your_path/datasets/waymo_open_organized/training_org training_org
ln -sfn your_path/datasets/waymo_open_organized/validation_org validation_orgThis will result in the following directory structure:
./code
├── data
│ └── waymo
│ ├── ImageSets
│ ├── raw_data
│ │ ├── training
│ │ │ ├── segment-10017090168044687777_6380_000_6400_000_with_camera_labels.tfrecord
│ │ │ └── segment-10023947602400723454_1120_000_1140_000_with_camera_labels.tfrecord
│ │ │
│ │ └── validation
│ │ ├── segment-10203656353524179475_7625_000_7645_000_with_camera_labels.tfrecord
│ │ └── segment-1024360143612057520_3580_000_3600_000_with_camera_labels.tfrecord
│ │
│ ├── training_org
│ │ ├── segment_id
│ │ ├── calib
│ │ ├── image_0
│ │ ├── image_1
│ │ ├── image_2
│ │ ├── image_3
│ │ ├── image_4
│ │ ├── label_0
│ │ ├── label_1
│ │ ├── label_2
│ │ ├── label_3
│ │ ├── label_4
│ │ ├── label_all
│ │ ├── projected_points_0
│ │ └── velodyne
│ │
│ └── validation_org
│ ├── segment_id
│ ├── calib
│ ├── image_0
│ ├── image_1
│ ├── image_2
│ ├── image_3
│ ├── image_4
│ ├── label_0
│ ├── label_1
│ ├── label_2
│ ├── label_3
│ ├── label_4
│ ├── label_all
│ ├── projected_points_0
│ └── velodyne
As a sanity check, you should see a total of more than 39k calib, image_0 and label_0 files in all the segment_ids of validation_org folder.
cd validation_org
find */calib -type f | wc -l
find */image_0 -type f | wc -l
find */label_0 -type f | wc -l
cd ..Finally, run the script setup_split.py to convert the segments into standard KITTI format and generate training and validation folders of the Waymo Val split:
python setup_split.pyThe script uses soft-links for efficient storage and creates the desired directory structure, which can be used by any KITTI based detector.
./code
├── data
│ └── waymo
│ ├── ImageSets
│ ├── raw_data
│ │ ├── training
│ │ └── validation
│ │
│ ├── training_org
│ ├── validation_org
│ │
│ ├── training
│ │ ├── calib
│ │ ├── image
│ │ └── label
│ │
│ └── validation
│ ├── calib
│ ├── image
│ └── label
Also prepare a small val split for testing:
cd ImageSets
sort -R --random-source=<(yes 123) val.txt | head -n 1000 > val_small.txt
cd ../../../Alter Way
We also upload the calib and label subfolders of the waymo training and validation split at this drive link.
Unzip the above file and place them as follows:
DEVIANT
├── data
│ └── waymo
│ ├── ImageSets
│ ├── training
│ │ ├── calib
│ │ └── label
│ │
│ └── validation
│ ├── calib
│ └── label
Then consider copying the corresponding images in the image sub-folder of training and validation folders to complete the folder structure:
DEVIANT
├── data
│ └── waymo
│ ├── ImageSets
│ ├── training
│ │ ├── calib
│ │ ├── image
│ │ └── label
│ │
│ └── validation
│ ├── calib
│ ├── image
│ └── label
Transfer
If you have Waymo dataset prepared and you need to transfer to your server, type the following:
cd your_desktop_waymo_location
cd training_org
rsync -qavR */calib/ [email protected]:/mnt/gs21/scratch/abhinavkumar/data/waymo_open_organized/training_org
rsync -qavR */image_0/ [email protected]:/mnt/gs21/scratch/abhinavkumar/data/waymo_open_organized/training_org
rsync -qavR */label_0/ [email protected]:/mnt/gs21/scratch/abhinavkumar/data/waymo_open_organized/training_org
cd ../validation_org
rsync -qavR */calib/ [email protected]:/mnt/gs21/scratch/abhinavkumar/data/waymo_open_organized/validation_org
rsync -qavR */image_0/ [email protected]:/mnt/gs21/scratch/abhinavkumar/data/waymo_open_organized/validation_org
rsync -qavR */label_0/ [email protected]:/mnt/gs21/scratch/abhinavkumar/data/waymo_open_organized/validation_org