| --- |
| title: WarehouseEye |
| emoji: 📦 |
| colorFrom: yellow |
| colorTo: blue |
| sdk: streamlit |
| sdk_version: 1.41.0 |
| app_file: frontend/app_space.py |
| pinned: false |
| license: mit |
| short_description: Industrial video intelligence on AMD MI300X |
| tags: |
| - amd-hackathon |
| - amd-mi300x |
| - qwen3-vl |
| - vision-language |
| - object-tracking |
| - warehouse-analytics |
| - on-premise |
| - open-source |
| --- |
| |
| # WarehouseEye |
|
|
| Operational intelligence for warehouse CCTV, built as an open-source pipeline and benchmarked on AMD MI300X. |
|
|
| This Space is a **CPU-friendly pre-rendered demo** designed for public exploration. It shows what the system can do while keeping the runtime lightweight for free-tier hosting. |
|
|
| ## Why this demo exists |
|
|
| Warehouse teams often need to audit long CCTV footage to answer operational and safety questions. Manual review is expensive, slow, and inconsistent. |
|
|
| WarehouseEye converts raw footage into structured timelines with: |
|
|
| - person tracking |
| - crop-level vision-language interpretation |
| - searchable event summaries |
| - benchmark and cost reporting |
|
|
| This Space exposes those results through a Streamlit UI with three tabs: |
|
|
| - **Operation Overview** |
| - **Ask the Video** |
| - **Performance Dashboard** |
|
|
| ## What runs here vs full system |
|
|
| This Space intentionally runs in a constrained mode: |
|
|
| - no FastAPI backend calls |
| - no live vLLM inference |
| - no model hosting in the Space |
|
|
| Instead, it reads pre-rendered outputs from `data/prerendered/<video_id>/`: |
|
|
| - `timeline.json` |
| - `benchmarks.json` |
| - `videos/input_video.mp4` |
| - selected track crops |
|
|
| For free-form natural-language queries and live inference, deploy the full repository locally or on AMD GPU infrastructure. |
|
|
| ## AMD MI300X relevance |
|
|
| WarehouseEye is tuned to showcase production-minded multimodal workloads on AMD hardware. MI300X provides 192 GB HBM, which is especially useful for single-node deployment of larger vision-language systems without coordinating multi-GPU sharding. |
|
|
| The performance dashboard in this Space reports **real measured benchmark outputs** captured from MI300X runs and stored in the pre-rendered artifacts. |
|
|
| ## Links |
|
|
| - GitHub: [https://github.com/Davidachoy/warehouseeye](https://github.com/Davidachoy/warehouseeye) |
| - Hugging Face Space: [https://huggingface.co/spaces/lablab-ai-amd-developer-hackathon/warehouseeye](https://huggingface.co/spaces/lablab-ai-amd-developer-hackathon/warehouseeye) |
| - Author X/Twitter: [https://x.com/achoy__](https://x.com/achoy__) |
|
|
| ## Disclaimer |
|
|
| Demo running on CPU with pre-rendered results. Real-time inference requires AMD MI300X. |
