levirdetnet / README.md
multimodalart's picture
multimodalart HF Staff
Upload folder using huggingface_hub
fee9d25 verified
|
Raw History Blame Contribute Delete
2.1 kB

A newer version of the Gradio SDK is available: 6.32.0

Upgrade
metadata
title: LEVIRDetNet
emoji: 🛰️
colorFrom: indigo
colorTo: purple
sdk: gradio
sdk_version: 6.31.0
app_file: app.py
pinned: false
license: other
short_description: 159-category aerial object detection with LEVIRDetNet
python_version: '3.10'
startup_duration_timeout: 45m

LEVIRDetNet — Universal Remote Sensing Object Detection

Demo of LEVIRDetNet-159, the foundation detector from LEVIRDet: A Million-Scale 159-Category Dataset and Foundation Model for Universal Remote Sensing Object Detection (code).

Upload an aerial / satellite image and the model returns bounding boxes over 159 categories — a hierarchy spanning 30 base classes (plane, ship, vehicle, storage tank, harbor …) down to fine types (A220, Boeing-737, Airliner, Small_leisure_craft, warship classes such as CV / DD / LCS …).

What the demo does

  • Loads the complete yangqinzhe/LEVIRDetNet-159 checkpoint (DINOv3 ViT-S/16 backbone + RTDETR hybrid encoder + 6-layer DEIMv2 decoder + hierarchy-aware head + online GSD predictor) into GPU memory.
  • Resizes the input to 1024 × 1024 (keep_ratio=False), one image per batch, and draws predicted boxes in original-image coordinates with the mmdet DetLocalVisualizer — the same pipeline as the authors' levir_image_demo.py.
  • Two numerical guards from the reference implementation are replicated: a neutral zero prior for single-image batches and the inactive-query reference-point NaN repair after query grouping.

Example images

The bundled samples come from the LEVIRDet release repository (assets/demo_samples/), released alongside the dataset, code and checkpoints.

Model card

  • Checkpoint: yangqinzhe/LEVIRDetNet-159 (best_coco_weighted_bbox_mAP_epoch_110_159class.pth)
  • Architecture: DEIMV2GSDGuided — DEIMv2 with DINOv3 ViT-S/16 (dinov3_vits16plus), GSD-conditioned FiLM query modulation, hierarchical prediction head (path-score collapse, no NMS).