Rewatiramans/bytetrack-onnx-static

ByteTrack ONNX Models

ONNX exports of pretrained ByteTrack models from the official ByteTrack repository.

Original repository:

• https://github.com/FoundationVision/ByteTrack

Included Models

Model	Description
bytetrack_ablation.onnx	Ablation model from paper
bytetrack_x_mot17.onnx	Highest accuracy MOT17 model
bytetrack_l_mot17.onnx	Large model
bytetrack_m_mot17.onnx	Medium model
bytetrack_s_mot17.onnx	Small production-friendly model
bytetrack_x_mot20.onnx	MOT20 model

Some models may also include:

model.onnx.data

This file contains external tensor weights for large ONNX exports. Keep it beside the .onnx file.

---

Export Notes

These models were exported using:

• PyTorch
• Official ByteTrack export script
• ONNX Runtime compatible format

The original ByteTrack repository required small compatibility fixes for modern PyTorch versions.

---

Input Shape

These exports use static input shapes.

Example:

[1, 3, 608, 1088]

Meaning:

• batch size = 1
• channels = 3 (RGB)
• height = 608
• width = 1088

Your input must exactly match the exported shape unless you re-export with dynamic axes.

---

ONNX Runtime Inference Example

import onnxruntime as ort
import numpy as np

session = ort.InferenceSession(
    "bytetrack_s_mot17.onnx",
    providers=["CPUExecutionProvider"]
)

img = np.random.rand(
    1,
    3,
    608,
    1088
).astype(np.float32)

outputs = session.run(
    None,
    {"images": img}
)

print(outputs[0].shape)

Expected output:

(1, 13566, 6)

---

TensorRT

These models can be converted to TensorRT engines using:

• trtexec
• TensorRT Python API
• DeepStream

Example:

trtexec \
--onnx=bytetrack_s_mot17.onnx \
--saveEngine=bytetrack_s.engine

---

Production Notes

Recommended models:

Scenario	Recommended Model
Maximum accuracy	bytetrack_x_mot17
Balanced production deployment	bytetrack_m_mot17
Lightweight deployment	bytetrack_s_mot17

The s model is generally the best balance between:

• speed
• VRAM usage
• latency
• tracking quality

---

Important

ByteTrack itself is only the tracker.

These ONNX models contain the detector component used alongside ByteTrack tracking logic.

Typical pipeline:

video frame
→ detector inference
→ ByteTrack association/tracking
→ tracked objects

---

Credits

ByteTrack Paper:

@article{zhang2022bytetrack,
  title={ByteTrack: Multi-Object Tracking by Associating Every Detection Box},
  author={Zhang, Yifu and Sun, Peize and Jiang, Yi and Yu, Dongdong and Weng, Fucheng and Yuan, Zehuan and Luo, Ping and Liu, Wenyu and Wang, Xinggang},
  booktitle={ECCV},
  year={2022}
}

Official repository:

• https://github.com/FoundationVision/ByteTrack