Immortal_tracker for CTRL

This is a modified unofficial version of ImmortalTracker, tailored to easier use for CTRL. We make the follow modifications:

Calculate 3D IoU on GPU for much faster tracking, requiring MMDet3D (<v1.0).
Support saving the extended boxes (by motion) for the forward tracking part in CTRL.
Improve readability.
Automatically evaluation on WOD, requiring Waymo evaluation tools.

These modifications is showcased in the config file: configs/immortal_for_ctrl_keep10.yaml. Users could compare it with official config to understand the differences.

Please follow the official instruction of ImmortalTracker to prepare the data, and then user could run run_mot.sh to get the tracking results. Note you may need to rename some variables in run_mot.sh to fit your needs.

Feel free to open an issue or contact me ([email protected]) if you have any problem.

Original README as follows

Prerequisite

Our code is tested for Python 3.6.
To install required liabraries:

pip install -r requirements.txt

Waymo Open Dataset

Prepare dataset & off-the-shelf detections

Download WOD perception dataset:

#Waymo Dataset         
└── waymo
       ├── training (not required)  
       ├── validation   
       ├── testing

To extract timestamp infos/ego infos from .tfrecord files, run the following:

bash preparedata/waymo/waymo_preparedata.sh  /<path to WOD>/waymo

Run the following to convert detection results into to .npz files. The detection results should be in official WOD submission format(.bin)
We recommand you to use CenterPoint(two-frame model for tracking) detection results for reproducing our results. Please follow https://github.com/tianweiy/CenterPoint or email its author for CenterPoint detection results.

bash preparedata/waymo/waymo_convert_detection.sh <path to detection results>/detection_result.bin cp

#you can also use other detections:
#bash preparedata/waymo/waymo_convert_detection.sh <path to detection results> <detection name>

Inference

Use the following command to start inferencing on WOD. The validation set is used by default.

python main_waymo.py --name immortal --det_name cp --config_path configs/waymo_configs/immortal.yaml --process 8

Evaluation with WOD official devkit:

Follow https://github.com/waymo-research/waymo-open-dataset to build the evaluation tools and run the following command for evaluation:

#Convert the tracking results into .bin file
python evaluation/waymo/pred_bin.py --name immortal
#For evaluation
<path to WOD devkit>/bazel-bin/waymo_open_dataset/metrics/tools/compute_tracking_metrics_main mot_results/waymo/validation/immortal/bin/pred.bin <path to WOD tracking ground truth file>/validation_gt.bin

nuScenes Dataset

Prepare dataset & off-the-shelf detections

Download nuScenes perception dataset

# For nuScenes Dataset         
└── NUSCENES_DATASET_ROOT
       ├── samples       
       ├── sweeps       
       ├── maps         
       ├── v1.0-trainval 
       ├── v1.0-test

To extract timestamp infos/ego infos, run the following:

bash preparedata/nuscenes/nu_preparedata.sh <path to nuScenes>/nuscenes

Run the following to convert detection results into to .npz files. The detection results should be in official nuScenes submission format(.json)
We recommand you to use centerpoint(two-frame model for tracking) detection results for reproducing our results.

bash preparedata/nuscenes/nu_convert_detection.sh  <path to detection results>/detection_result.json cp

#you can also use other detections:
#bash preparedata/nuscenes/nu_convert_detection.sh <path to detection results> <detection name>

Inference

Use the following command to start inferencing on nuScenes. The validation set is used by default.

python main_nuscenes.py --name immortal --det_name cp --config_path configs/nu_configs/immortal.yaml --process 8

Evaluation with nuScenes official devkit:

Follow https://github.com/nutonomy/nuscenes-devkit to build the official evaluation tools for nuScenes. Run the following command for evaluation:

#To convert tracking results into .json format
bash evaluation/nuscenes/pipeline.sh immortal
#To evaluate
python <path to nuscenes-devkit>/nuscenes-devkit/python-sdk/nuscenes/eval/tracking/evaluate.py \
"./mot_results/nuscenes/validation_2hz/immortal/results/results.json" \
--output_dir "./mot_results/nuscenes/validation_2hz/immortal/results" \
--eval_set "val" \
--dataroot <path to nuScenes>/nuscenes

regatte / immortaltracker-for-ctrl Goto Github PK