Detection Quality Matters

Tracking quality starts at the detector. If it misses an object, the tracker never gets a chance. This guide isolates the effect of detector choice by running ByteTrack with three models of increasing accuracy on the MOT17 benchmark.

What you'll learn:

Run the same tracker with different detection models
Measure how detector choice impacts tracking metrics
Compare YOLO26 Nano, RF-DETR Nano, and RF-DETR Medium on MOT17

Install

Install trackers with the detection extra to enable built-in model support.

pip install "trackers[detection]"

For more options, see the install guide.

Detection Models

We pick three models that span a wide accuracy range on COCO, from a lightweight YOLO to a mid-size transformer detector. The gap in COCO accuracy between the weakest and strongest model is nearly 15 AP. The question is how much of that carries over to tracking.

RF-DETR vs top object detectors on MS COCO

Comparison of RF-DETR against other top real-time detectors on MS COCO.

Model	COCO AP50	COCO AP50:95	Latency (ms)
YOLO26 Nano	55.8	40.3	1.7
RF-DETR Nano	67.6	48.4	2.3
RF-DETR Medium	73.6	54.7	4.4

Download Data

Download the MOT17 validation split (frames for detection, annotations for evaluation) — see the download guide for the full command and available options.

Run the Experiment

Run ByteTrack with default parameters three times, changing only the detection model each time.

YOLO26 Nano

Single sequenceAll sequences

trackers track \
    --source ./data/mot17/val/MOT17-13-FRCNN/img1 \
    --detection.model yolo26n-640 \
    --tracker bytetrack \
    --filters.classes [person] \
    --output.mot_results results/yolo26n/MOT17-13-FRCNN.txt

for seq in MOT17-02-FRCNN MOT17-04-FRCNN MOT17-05-FRCNN MOT17-09-FRCNN MOT17-10-FRCNN MOT17-11-FRCNN MOT17-13-FRCNN; do
    trackers track \
        --source ./data/mot17/val/$seq/img1 \
        --detection.model yolo26n-640 \
        --tracker bytetrack \
        --filters.classes [person] \
        --output.mot_results results/yolo26n/$seq.txt
done

ByteTrack with YOLO26 Nano on MOT17-13.

RF-DETR Nano

Single sequenceAll sequences

trackers track \
    --source ./data/mot17/val/MOT17-13-FRCNN/img1 \
    --detection.model rfdetr-nano \
    --tracker bytetrack \
    --filters.classes [person] \
    --output.mot_results results/rfdetr-nano/MOT17-13-FRCNN.txt

for seq in MOT17-02-FRCNN MOT17-04-FRCNN MOT17-05-FRCNN MOT17-09-FRCNN MOT17-10-FRCNN MOT17-11-FRCNN MOT17-13-FRCNN; do
    trackers track \
        --source ./data/mot17/val/$seq/img1 \
        --detection.model rfdetr-nano \
        --tracker bytetrack \
        --filters.classes [person] \
        --output.mot_results results/rfdetr-nano/$seq.txt
done

ByteTrack with RF-DETR Nano on MOT17-13.

RF-DETR Medium

Single sequenceAll sequences

trackers track \
    --source ./data/mot17/val/MOT17-13-FRCNN/img1 \
    --detection.model rfdetr-medium \
    --tracker bytetrack \
    --filters.classes [person] \
    --output.mot_results results/rfdetr-medium/MOT17-13-FRCNN.txt

for seq in MOT17-02-FRCNN MOT17-04-FRCNN MOT17-05-FRCNN MOT17-09-FRCNN MOT17-10-FRCNN MOT17-11-FRCNN MOT17-13-FRCNN; do
    trackers track \
        --source ./data/mot17/val/$seq/img1 \
        --detection.model rfdetr-medium \
        --tracker bytetrack \
        --filters.classes [person] \
        --output.mot_results results/rfdetr-medium/$seq.txt
done

ByteTrack with RF-DETR Medium on MOT17-13.

Evaluate

Evaluate each run against ground truth using CLEAR, HOTA, and Identity metrics.

YOLO26 Nano

trackers eval \
    --gt_dir ./data/mot17/val \
    --predictions_dir results/yolo26n \
    --metrics '[CLEAR,HOTA,Identity]' \
    --columns '[MOTA,HOTA,IDF1]'

Output:

                                MOTA    HOTA    IDF1
----------------------------------------------------
COMBINED                      23.444  32.874  34.411

RF-DETR Nano

trackers eval \
    --gt_dir ./data/mot17/val \
    --predictions_dir results/rfdetr-nano \
    --metrics '[CLEAR,HOTA,Identity]' \
    --columns '[MOTA,HOTA,IDF1]'

Output:

                                MOTA    HOTA    IDF1
----------------------------------------------------
COMBINED                      25.667  35.735  38.182

RF-DETR Medium

trackers eval \
    --gt_dir ./data/mot17/val \
    --predictions_dir results/rfdetr-medium \
    --metrics '[CLEAR,HOTA,Identity]' \
    --columns '[MOTA,HOTA,IDF1]'

Output:

                                MOTA    HOTA    IDF1
----------------------------------------------------
COMBINED                      29.141  38.637  41.950

Results

Same tracker, same data, same parameters. The only difference is the detector.

Detector	MOTA	HOTA	IDF1
YOLO26 Nano	23.444	32.874	34.411
RF-DETR Nano	25.667	35.735	38.182
RF-DETR Medium	29.141	38.637	41.950

RF-DETR Medium leads across every metric, showing that a stronger detector directly lifts tracking quality.

Takeaway

Before tweaking tracker hyperparameters, invest in detection quality. The results above show that swapping the detector alone produces larger gains than most tracker-level optimizations.