Transformer-based decision network for multiple object tracking (TADN)

Check our final paper in CVIU journal here
or check our pre-print publication here

Highlights

• Optimization during inference
• TADN can directly infer assignment pairs between detections and active targets in a single forward pass of the network
• TADN is integrated in a rather simple MOT framework
• Coupled with a novel training strategy for efficient end-to-end training

Benchmark results

Dataset	Detections	MOTA	IDF1	MT	ML	FP	FN	IDSW	Frag	Hz
MOT17	Public	54.6	49.0	22.4	30.2	36285	214857	4869	7821	10.0
MOT17	Private	69.0	60.8	45.7	13.6	47466	124623	2955	4119	-
MOT20	Private	68.7	61.0	57.4	14.3	27135	133045	1707	2321	-
UA-DETRAC	Private	23.7	-	61.2	8.2	31417	198714	198714	-	-

Notes

• TADN performance for MOT17 in Hz is achieved using a NVIDIA Geforce GTX2080Ti. Performance may vary for different hardware configurations.
• TADN is trained exclusively on each benchmark's provided training data.
• MOT17 and MOT20 private detections results are achieved using a community pretrained YOLO-X detector publicly available here.
• UA-DETRAC metrics are the CLEAR metrics along the detector's PR-curve.

Installation

To install tadn-mot you must first install its dependencies as found in requirements.txt. For ease, you can use docker.

Docker

To use tadn-mot with docker, first you must build the appropriate docker image

docker build . -t tadn-mot

To start a TADN-MOT container in interactive mode:

docker run -it [GPU_RELATED_OPTS] [MISC OPTIONS] tadn-mot bash

Usage

TADN is configured for training in either MOT17 or UA-DETRAC benchmarks. To speed up thetraining process and allow for quick experimentation many inputs can be precomputed. However, TADN can be deployed to provide real-time tracking with minimal effort.

Prepare Data

• MOT 17: Download dataset from this link and unzip.
• UA-DETRAC: Download dataset from this link and unzip.

Note : For MOT17 benchmark detections from FRCNN, SDP, and DPM are provided. For UA-DETRAC detections from CompACT, RCNN, DPM and ACF detectors are provided. To use EB detections download detections from here.

Note 2 : To enable validation for UA-DETRAC using the trackeval repo, you must firts run the following convenience script:

python -m tadn.scripts.detrac_generate_MOTC_gt PATH_TO_DATASET_ROOT --dset_mode "train"
python -m tadn.scripts.detrac_generate_MOTC_gt PATH_TO_DATASET_ROOT --dset_mode "test"

Choose appearance features CNN encoder:

1. ImageNet pretrained ResNet-18
   • No further actions needed
   • Mediocre performance
2. Re-id pretrained ResNet-50
   • Use these instructions to pretrain the model on MOT17
   • Better performance (~ 5%)

Precompute appearance vectors:

• MOT17 & Re-id CNN features

python -m tadn.scripts.precompute_appearance_vectors PATH_TO_DATASET_ROOT --dset_type mot-challenge --dset_version MOT17 --feature_extractor reid --reid_ckpt PATH_TO_REID_CHECKPOINT

• MOT17 & Resnet-18 CNN features

python -m tadn.scripts.precompute_appearance_vectors PATH_TO_DATASET_ROOT --dset_type mot-challenge --dset_version MOT17 --feature_extractor resnet18

• UA-DETRAC & Re-id CNN features

python -m tadn.scripts.precompute_appearance_vectors PATH_TO_DATASET_ROOT --dset_type detrac --detector EB --feature_extractor reid --reid_ckpt PATH_TO_REID_CHECKPOINT

• MOT17 & Resnet-18 CNN features

python -m tadn.scripts.precompute_appearance_vectors PATH_TO_DATASET_ROOT --dset_type mot-challenge --dset_version MOT17 --feature_extractor resnet18

Camera Motion Compensation

If you intend to use CMC in your model, you must first precompute frame to frame affine transforms for the target benchmark using the ECC method.

• MOT17

python -m tadn.scripts.precompute_ecc PATH_TO_DATASET_ROOT --dset_type mot-challenge --dset_mode "train" --dset_version MOT17
python -m tadn.scripts.precompute_ecc PATH_TO_DATASET_ROOT --dset_type mot-challenge --dset_mode "val" --dset_version MOT17

• UA-DETRAC

python -m tadn.scripts.precompute_ecc PATH_TO_DATASET_ROOT --dset_type detrac --dset_mode "train"
python -m tadn.scripts.precompute_ecc PATH_TO_DATASET_ROOT --dset_type detrac --dset_mode "test"

Train a TADN model

This repository uses a JSON based configuration system for defining:

• Model architecture
• Training parameters
• Dataset configuration
• Data input pipeline
• Logger configuration
• Miscelaneous hyperparameters

Note : JSON config examples can be found in the sample_configs directory

Train on MOT17 or UA-DETRAC:

python -m tadn.online_training PATH_TO_JSON_CONFIG

Inference

Inference with a pre-trained TADN module supports the MOTChallenge format for the output results.

To perform inference on the val/test subsets:

python -m tadn.scripts.inference PATH_TO_CKPT PATH_TO_JSON_CONFIG 

To perform inference on the whole dataset:

python -m tadn.scripts.inference PATH_TO_CKPT PATH_TO_JSON_CONFIG --inference_train

Evaluation

To evaluate a pre-trained TADN model on either MOT17 or UA-DETRAC datasets:

python -m tadn.scripts.validate PATH_TO_CKPT PATH_TO_JSON_CONFIG

Note : Which dataset you evaluate is configured in the JSON config

Note 2: For MOT17, test set is unavailable. You can use "split": "half" to perform evaluation on a 50/50 split of MOT17 dataset. Similarly you can train a model on that split to replicate ablation study experiments. To evaluate on the official MOT17 test set, please use the "inference" script to submit inference results to the official evaluation server.

Validate training strategy

To estimate expected tracking performance when training TADN using the online training strategy you can run:

python -m tadn.scripts.validate_LAM PATH_TO_JSON_CONFIG

Pre-trained TADN models

Download pretrained models and their json configs for MOT17 and UA-DETRAC from zenodo !

Cite us!

If you use TADN in your research or wish to refer to the baseline results published here, please use the following BibTeX entry:

@article{psalta2024transformer,
 title={Transformer-based assignment decision network for multiple object tracking},
 author={Psalta, Athena and Tsironis, Vasileios and Karantzalos, Konstantinos},
 journal={Computer Vision and Image Understanding},
 pages={103957},
 year={2024},
 publisher={Elsevier}
}