README.md

---
license: mit
tags:
- spatial-temporal-memory
- video-understanding
- object-tracking
- depth-estimation
---

# Spatial-Temporal Memory System - Model Weights

This repository contains pre-trained model weights for the [Spatial-Temporal Memory System](https://github.com/Sealical/spatial-temporal-memory).

## Directory Structure

```
weights/
├── Depth-Anything-V2/      # Depth-Anything-V2 checkpoints
├── DUSt3R/                 # DUSt3R pose / reconstruction checkpoints
├── Fast-ReID/              # Re-identification checkpoints
├── GroundingDINO/          # Open-vocabulary detection checkpoints
├── Metric3D/               # Metric depth checkpoints
├── Qwen3-Embedding-8B/     # Large text embedding model
├── RAFT/                   # Optical flow checkpoints
├── SAM2/                   # SAM2 segmentation checkpoints
├── checkpoints/            # ViPE / auxiliary checkpoints
├── dinov2/                 # DINOv2 local weights
├── droid_slam/             # DROID-SLAM related weights
├── geocalib/               # Geo calibration weights
├── sam3/                   # SAM3 local bundle
├── osnet_x0_25_msmt17.pt   # Lightweight ReID checkpoint
└── yolov8n.pt              # Lightweight YOLO checkpoint
```

## Usage

Download weights using the provided script:

```bash
bash scripts/download_weights_from_hf.sh
```

Or manually with huggingface-cli:

```bash
huggingface-cli download sealical/spatial-temporal-mem-models --local-dir weights/
```

## Notes

- This mirror follows the project's current `weights/` layout.
- Some folders contain very large checkpoints and model bundles intended for local sync rather than GitHub storage.
- If you only need a smaller public subset, use `bash scripts/download_weights.sh`.

## License

Model weights are subject to their original licenses. Please check individual model repositories for details.