FlashTrack

Ultra-lightweight real-time multi-object tracking built on ShuffleNetV2 backbone with DeepSORT/ByteTrack-style tracking. Part of the FlashVision family (alongside FlashDet, FlashSeg, FlashOCR).

What is FlashTrack?

FlashTrack combines a lightweight ReID (re-identification) feature extractor with Kalman filtering for robust multi-object tracking. It's designed for edge deployment — small model size, fast inference, and strong tracking accuracy.

Key features:

• Tiny models — 0.3M to 1.5M parameters (< 3 MB FP16)
• Three tracker backends — ByteTracker, SORTTracker, DeepSORTTracker
• ReID-powered — Deep appearance features for robust re-identification
• LoRA fine-tuning — 6 variants for parameter-efficient adaptation
• Knowledge Distillation — Train small models from large teachers
• ONNX export — Deploy anywhere

Model Zoo

Model	Backbone	ReID Dim	Params	FP16 Size
FlashTrack-m-0.5x	ShuffleNetV2 0.5x	128	~0.3M	~0.6 MB
FlashTrack-m	ShuffleNetV2 1.0x	128	~0.8M	~1.6 MB
FlashTrack-m-1.5x	ShuffleNetV2 1.5x	256	~1.5M	~3.0 MB

Trackers

Tracker	Description	ReID?	Speed
ByteTracker	Two-stage IoU association (ByteTrack)	No	Fastest
SORTTracker	Kalman + Hungarian + IoU	No	Fast
DeepSORTTracker	Kalman + ReID cosine distance	Yes	Accurate

Installation

# From source
git clone https://github.com/FlashVision/FlashTrack.git
cd FlashTrack
pip install -e ".[all]"

# Verify installation
flashtrack check

Quick Start

Python API

from flashtrack import FlashTracker, ByteTracker, Predictor

# Run tracking on video
predictor = Predictor(
    model_path="workspace/model_best_inference.pth",
    tracker_type="bytetrack",
)
predictor.track_video("input.mp4", output_dir="output/")

# Use ByteTracker directly
tracker = ByteTracker(track_thresh=0.5, track_buffer=30)
tracks = tracker.update(detections)  # [N, 5] -> [M, 7]

CLI

# Train ReID model
flashtrack train --train-data data/MOT17/train --model-size m --epochs 120

# Run tracking
flashtrack track --source video.mp4 --tracker bytetrack

# Export to ONNX
flashtrack export --model workspace/model_best_inference.pth --simplify

Training

Standard Training

from flashtrack import Trainer

trainer = Trainer(
    model_size="m",
    epochs=120,
    batch_size=64,
    train_data="data/MOT17/train",
    val_data="data/MOT17/val",
    amp=True,
)
trainer.train()

LoRA Fine-Tuning

trainer = Trainer(
    model_size="m",
    lora=True,
    lora_rank=8,
    lora_variant="standard",  # or "dora", "lora_plus", "adalora", "ortho", "lora_fa"
    train_data="data/MOT17/train",
)
trainer.train()

Knowledge Distillation

Train a smaller student from a larger teacher:

flashtrack train --config configs/flashtrack_m_kd.yaml

Architecture

Input Crop [B, 3, 128, 64]
    │
    ▼
ShuffleNetV2 Backbone
    │
    ▼
Feature Encoder (1x1 + DW 3x3 + 1x1 → GAP)
    │
    ▼
ReID Head (BN → FC → BN Neck)
    │
    ├── Embedding [B, 128] (inference)
    └── Logits [B, num_ids] (training)

YAML Configuration

model:
  backbone_size: "1.0x"
  reid_dim: 128
  num_ids: 750
  input_size: [128, 64]

data:
  train_images: data/MOT17/train
  val_images: data/MOT17/val

train:
  epochs: 120
  batch_size: 64
  learning_rate: 0.0003

License

MIT License — free for commercial and academic use.