Tiny-object detectors stop guessing query counts: BUTR sets the budget from encoder confidence, reaching 35.7 AP with 316 queries on AI-TOD-v2
Synopsis
In dense tiny-object detection, dynamic query designs previously entangled how many queries are decoded with which queries are kept, obscuring the source of gains. BUTR reframes dynamic query as allocating decoder query budget by object density: it sums encoder proposal confidences into a density proxy and maps it linearly to a per-image query count, keeping standard Top-k initialization without learned modules or heuristic filtering. On AI-TOD-v2, BUTR-L reaches 35.7 AP with 316 queries on average, versus 35.2 AP and 600 queries for same-scale D-FINE-L.
Interpretation
In dense tiny-object detection, the decoder query count now varies per image with object density instead of using one fixed value for every image. A fixed query budget is efficient for most sparse aerial images but may be insufficient for the rare crowded scenes with hundreds of instances; uniformly raising it adds overhead and more false positives on sparse images. On AI-TOD-v2, BUTR-L reaches 35.7 AP versus 35.2 AP for same-scale fixed-query D-FINE-L, while BUTR-L uses 316 queries on average and D-FINE-L uses 600.
The query budget is computed directly from encoder proposal confidences already present, without an extra learned predictor or filtering pipeline. Earlier dynamic-query methods either predicted a density map and used a classifier to pick a preset budget bin, or used a learned window mask plus heuristic filtering to retain queries. The budgeter sums each proposal's maximum sigmoid class confidence as a density proxy and multiplies it by a safety factor; this proxy correlates with true object count at Pearson 0.86 and Spearman 0.85 on AI-TOD-v2, and 0.82 and 0.81 on VisDrone.
On the milder VisDrone benchmark, the same budgeting rule still yields a small accuracy gain. VisDrone has a weaker density tail than AI-TOD-v2, so the budgeting view predicts smaller dynamic-query gains. BUTR-L reaches 29.8 AP on VisDrone versus 29.6 AP for same-scale D-FINE-L, with 398 queries on average for BUTR-L and 400 for D-FINE-L.
Perspective
The demonstrated scope is two aerial tiny-object benchmarks, AI-TOD-v2 and VisDrone, plus a sorghum grain-counting dataset and an inference-time plug-in experiment on COCO; the intended readers are detection researchers and deployment engineers working on two-stage DETR query initialization.
The budgeting signal depends on the numerical scale of confidences, so score calibration, proposal count or domain changes may affect allocation even when rankings are unchanged; other dense settings such as crowded products, cells or particles remain to be evaluated; dynamic-budget inference incurs synchronization and engine-boundary overhead in the current TensorRT deployment, so fewer queries do not necessarily mean lower latency; the high-density subsets contain few images, 26 on AI-TOD-v2 and 7 on VisDrone, so their AP is sensitive to test-sample composition.