Skip to content

Releases: Geekgineer/motcpp

reid-models-v1.0.0

07 Feb 10:31

Choose a tag to compare

ReID Weight Files (.pt) for Appearance-Based Trackers

A collection of 33 ReID models adapted for appearance-feature-based multi-object trackers.

Adapted from:

These models are typically used in StrongSORT, DeepSORT variants, and other ReID-based MOT pipelines.


Model Families Overview

Model Family # Models Description
ResNet50 6 Strong baseline backbone
MLFN 3 Multi-Level Factorization Network
HACNN 3 Attention-based ReID model
MobileNetV2 6 Lightweight, real-time friendly
OSNet 15 Omni-scale, state-of-the-art

Total models: 33


Model Family Details

ResNet50 (6)

  • Classic CNN-based ReID backbone
  • Good balance between accuracy and speed
  • Common baseline for person ReID tasks

MLFN – Multi-Level Factorization Net (3)

  • Learns discriminative latent factors
  • Robust to pose, viewpoint, and appearance changes

HACNN – Harmonious Attention CNN (3)

  • Combines spatial and channel attention
  • Focused on fine-grained person discrimination

MobileNetV2 (6)

  • Optimized for low latency and efficiency
  • Suitable for real-time tracking and edge devices
  • Lower computational cost

OSNet – Omni-Scale Network (15)

Available variants:

  • OSNet x1.0
  • OSNet x0.75
  • OSNet x0.5
  • OSNet x0.25
  • OSNet-IBN (Instance-Batch Normalization)
  • OSNet-AIN (Adaptive Instance Normalization)

Key advantages:

  • Excellent generalization across datasets
  • Strong performance in multi-object tracking
  • Scalable architectures for speed/accuracy trade-offs

Typical Applications

  • Appearance feature extraction for MOT
  • StrongSORT / DeepSORT-based trackers
  • Person re-identification across cameras
  • Offline and real-time tracking pipelines
    **

benchmark-data-v1.0

22 Jan 10:19

Choose a tag to compare

✓ MOT17-mini.tar.gz (2.8M) - Test sequences
✓ yolox_dets.tar.gz (113M) - YOLOX detections
✓ reid_embs.tar.gz (110M) - ReID embeddings

These data are used for the CI Pipeline.