TrafficVision

TrafficVision trains and evaluates DETRAC vehicle detectors and runs four tracking-by-detection pipelines. main.py is the command-line entry point.

Models

• faster_rcnn: TorchVision Faster R-CNN ResNet-50-FPN with an RPN.
• fast_rcnn: shared image features and RoI Align with Selective Search boxes.
• rcnn: independent ResNet crop features for Selective Search boxes.

Fast R-CNN and R-CNN need opencv-contrib-python for cv2.ximgproc Selective Search. Region proposals are cached under outputs/selective_search_cache/. R-CNN is substantially slower because it runs the CNN separately on each box.

Setup

Install a PyTorch and TorchVision version pair that matches your operating system and accelerator using the official PyTorch installation selector. Then install the remaining dependencies:

pip install -r requirements.txt

The DETRAC dataset should be under dataset/DETRAC/. Training uses the 60 annotated train sequences and holds out whole sequences for validation. The 40 test sequences are reserved for final evaluation.

The default quick-training profile samples up to 50 evenly spaced frames from each training sequence and runs 3 epochs. This reduces the training workload from about 82,000 frames per epoch to about 3,000 sampled train/validation frames per epoch without deleting or copying the source dataset. It is intended to fit within roughly 4 hours on a Tesla T4; actual time depends on the runtime, image storage speed, and GPU load. Set max_frames_per_sequence to null and adjust epochs in configs/train.yaml to use the full dataset.

Commands

List the supported detector names:

python main.py models

Train a detector. --model may be faster_rcnn, fast_rcnn, or rcnn. Initialization can be pretrained or scratch.

python main.py train --model faster_rcnn --initialization pretrained
python main.py train --model faster_rcnn --initialization scratch
python main.py train --model fast_rcnn --initialization pretrained
python main.py train --model rcnn --initialization pretrained

Training resumes automatically from last.pth in the selected output directory when it exists. Each invocation runs the configured number of epochs; on resume, it counts those epochs from the saved epoch. Use --fresh to ignore an existing checkpoint and start a new run. Annotated frames whose image files are missing are skipped with a warning; the dataset root and sequence directory layout still need to be correct.

Run the four detector experiments, evaluate each one, then compare all trackers on the default Faster R-CNN pretrained detections:

bash run_experiments.sh

Evaluate a previously trained model. Select the same initialization used during training so the matching checkpoint is loaded.

python main.py evaluate --model faster_rcnn --initialization pretrained
python main.py evaluate --model faster_rcnn --initialization scratch
python main.py evaluate --model fast_rcnn --initialization pretrained
python main.py evaluate --model rcnn --initialization pretrained

Options such as epochs, batch size, device, validation split, score thresholds, and paths are in configs/train.yaml, configs/test.yaml, and configs/model.yaml. Command-line values override the corresponding config values. Training checkpoints are stored in directories named by model and initialization, such as outputs/detrac_faster_rcnn_scratch/. Evaluation writes to matching outputs/evaluation_<model>_<initialization>/ directories with per-sequence detection CSVs and metrics.json.

Tracking

First evaluate a trained detector on DETRAC Test. This exports per-sequence detection CSV files to outputs/evaluation_<model>_<initialization>/detections/. The tracking config points to the pretrained Faster R-CNN results by default. Then run one tracker against those detections:

python main.py track --tracker sort
python main.py track --tracker bytetrack
python main.py track --tracker botsort
python main.py track --tracker ocsort
python main.py track-all

The selected tracker can also be set in configs/tracking.yaml. Tracker settings include maximum lost age, minimum hits, IoU threshold and ByteTrack confidence thresholds. BoT-SORT uses optical-flow camera-motion compensation and color histograms as a lightweight appearance cue; it does not load a learned ReID model. Results are saved as per-sequence track CSV files and <tracker>/ subfolders under outputs/tracking/, so running multiple algorithms does not overwrite their CSVs. track-all also writes a comparison.json. Metrics include project-implemented HOTA, IDF1, MOTA and ID switches, reported per sequence and as an unweighted sequence macro average. DETRAC Test annotation IDs are used as ground truth.

Known gaps before benchmark or production use

• The tracker implementations are lightweight project versions, not exact reproductions of the authors' reference implementations. SORT uses a simple constant-velocity box model, OC-SORT does not implement every observation re-update step, and BoT-SORT uses color histograms instead of learned ReID.
• Tracking HOTA/IDF1/MOTA are implemented locally and should be cross-checked with the official TrackEval toolkit before reporting benchmark results.
• Training selects best.pth by validation loss. Validation AP/mAP-based checkpoint selection, resume-training, early stopping and a detailed epoch history are not implemented yet.
• There is no automated test suite or tracking visualization/export to video.
• Install a PyTorch/TorchVision pair compatible with your platform separately; requirements.txt lists the remaining packages. Put the DETRAC files under the documented dataset/DETRAC/ layout; datasets and model outputs are excluded from Git by .gitignore.

These gaps do not block running the current train → detection evaluation → tracking flow, but the tracker approximations and local metric implementation limit direct comparison with published benchmark numbers.