"""LEVIRDetNet: universal remote sensing object detection (159 categories). Runs the authors' LEVIRDetNet-159 checkpoint (DEIMv2 + DINOv3 ViT-S/16 backbone, GSD-guided queries, hierarchy-aware head) on an aerial/satellite image and draws the predicted bounding boxes. Mirrors the reference `levir_image_demo.py` path: 1024x1024 resize, single image per batch, hierarchy-collapsed scores, boxes in original-image coordinates. """ import os # ZeroGPU: import spaces before anything CUDA-touching. import spaces import contextlib import sys import types # mmcv-lite has no compiled CUDA ops. The vendored mmdet imports mmcv.ops # symbols at import time but never calls the CUDA kernels on this model's # inference path (decoder uses the pure-PyTorch deformable attention; the # head applies no NMS). Stub the ext loader so imports resolve. class _ExtStub(types.ModuleType): def __getattr__(self, name): if name.startswith('__') and name.endswith('__'): raise AttributeError(name) def _placeholder(*args, **kwargs): raise RuntimeError( f'mmcv CUDA op {name!r} called; not available in this no-ops build') return _placeholder sys.modules.setdefault('mmcv._ext', _ExtStub('mmcv._ext')) import glob import json import time from pathlib import Path import gradio as gr import numpy as np import torch APP_DIR = Path(__file__).resolve().parent sys.path.insert(0, str(APP_DIR)) from levir_net import build_model, load_image, singleton_prior, inactive_query_guard from mmcv.transforms import Compose from mmengine.dataset import pseudo_collate from mmengine.registry import init_default_scope from mmdet.registry import VISUALIZERS from mmdet.utils import register_all_modules MODEL_REPO = 'yangqinzhe/LEVIRDetNet-159' CKPT_NAME = 'best_coco_weighted_bbox_mAP_epoch_110_159class.pth' CONFIG_PATH = str(APP_DIR / 'configs/levirdetnet/levirdetnet-159class.py') register_all_modules(init_default_scope=True) init_default_scope('mmdet') print(f'[app] downloading checkpoint from {MODEL_REPO}/{CKPT_NAME} ...') from huggingface_hub import hf_hub_download CKPT_PATH = hf_hub_download(MODEL_REPO, CKPT_NAME) print(f'[app] checkpoint at {CKPT_PATH}') print('[app] building LEVIRDetNet-159 ...') MODEL, CFG, CLASSES = build_model(APP_DIR, CKPT_PATH, device='cuda') MODEL = MODEL.to('cuda') MODEL.eval() print(f'[app] model ready: {len(CLASSES)} classes') # Deterministic 159-class palette (mmdet 'random' palette seeds with 42). from mmdet.visualization.palette import get_palette PALETTE = get_palette('random', len(CLASSES)) MODEL.dataset_meta['palette'] = PALETTE SAMPLE_DIR = APP_DIR / 'samples' SAMPLE_IMAGES = sorted(SAMPLE_DIR.glob('*.jpg')) def _draw(image_path, labels, scores, boxes, score_thr): """Draw boxes with mmdet's own DetLocalVisualizer (authors' visualization path).""" from mmengine.structures import InstanceData from mmdet.structures import DetDataSample import cv2 img = cv2.imread(str(image_path)) img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) keep = [i for i, s in enumerate(scores) if s >= score_thr] instances = InstanceData() instances.bboxes = torch.tensor([boxes[i] for i in keep], dtype=torch.float32) instances.scores = torch.tensor([scores[i] for i in keep], dtype=torch.float32) instances.labels = torch.tensor([labels[i] for i in keep], dtype=torch.long) sample = DetDataSample() sample.pred_instances = instances sample.set_metainfo(dict(img_path=str(image_path), scale_factor=np.array([1.0, 1.0, 1.0, 1.0]))) vis = VISUALIZERS.build(dict(type='DetLocalVisualizer', name='levir', vis_backends=[])) vis.dataset_meta = dict(classes=CLASSES, palette=PALETTE) vis.set_image(img) vis.add_datasample('detection', img, sample, draw_gt=False, show=False, pred_score_thr=score_thr, out_file=None) return vis.get_image() @spaces.GPU(duration=30) def detect(image, score_threshold=0.3, max_boxes=300): """Detect objects in a remote sensing image with LEVIRDetNet. Args: image: input aerial/satellite image (any size; resized to 1024x1024 internally). score_threshold: minimum confidence score to draw/report a box. max_boxes: maximum detections drawn (highest-scoring kept). Returns: (annotated image, detections summary + full list as JSON) """ if image is None: raise gr.Error('Please upload an image first.') t0 = time.time() # Gradio hands us a numpy array (RGB). Save to a temp file so the mmdet # LoadImageFromFile pipeline handles it exactly like the reference demo. import tempfile with tempfile.NamedTemporaryFile(suffix='.jpg', delete=False) as f: # Gradio >= 4 numpy arrays are RGB; cv2.imwrite wants BGR. import cv2 cv2.imwrite(f.name, cv2.cvtColor(image, cv2.COLOR_RGB2BGR)) tmp_path = f.name try: data = load_image(tmp_path) batched = pseudo_collate([data]) with torch.no_grad(), singleton_prior(MODEL, torch), inactive_query_guard(MODEL, torch): outputs = MODEL.test_step(batched) sample = outputs[0] if isinstance(outputs[0], dict) else outputs[0] pred = sample['pred_instances'] if isinstance(sample, dict) else sample.pred_instances labels = pred.labels.detach().cpu().tolist() scores = pred.scores.detach().cpu().tolist() boxes = pred.bboxes.detach().cpu().tolist() vis_img = _draw(tmp_path, labels, scores, boxes, score_threshold) finally: with contextlib.suppress(Exception): os.unlink(tmp_path) keep = [i for i, s in enumerate(scores) if s >= score_threshold][:max_boxes] dets = [ dict(label=CLASSES[l], label_id=int(l), score=round(s, 4), bbox=[round(v, 1) for v in b]) for i, (l, s, b) in enumerate(zip(labels, scores, boxes)) if i in set(keep) ] elapsed = time.time() - t0 summary = dict(num_detections=len(dets), classes_present=sorted({d['label'] for d in dets}), score_threshold=score_threshold, elapsed_seconds=round(elapsed, 2)) return vis_img, json.dumps(dict(summary=summary, detections=dets), ensure_ascii=False, indent=2) CSS = ''' .dark .gradio-container, .dark { --bg: #0b0d12 !important; --button-primary-bg-hover: #6b7cff !important; } .gradio-container .prose h1 { font-size: 26px; } ''' with gr.Blocks() as demo: gr.HTML('''
159-category aerial detection with GSD-guided queries and a hierarchy-aware head (DEIMv2 + DINOv3 ViT-S/16). From the paper LEVIRDet: A Million-Scale 159-Category Dataset and Foundation Model for Universal Remote Sensing Object Detection.