cs.CVOct 3, 2026
SaveAny-scale Object Detection using Arbitrary-scaled Images
Organizations: Toyota Technological Institute
Abstract
This paper proposes any-scale object detection using arbitrary-scale super-resolution for continuously rescaling object images, while general multi-scale object detection uses discretely rescaled appearance representations. However, a naive usage of super-resolution produces many false-positive detections if many super-resolution images are independently fed into an object detector. Our method suppresses these false positives by predicting scale proposal maps, each of which represents a set of pixels appropriate for each super-resolution scale.
Figures & tables
Figure 1: Our proposed any-scale object detection shown in (b). Gray boxes in (a) indicate downsampled and upsampled images/features.
Figure 2: Our any-scale object detection (b), consisting of a single pair of the arbitrary-scale region proposal (ASRP) and ASSR networks, is executed multiple times with any . But, multiple region proposals and SR networks are used independently in (a). ASRP is detailed in Fig. 3 .
Figure 3: Detail of the proposed ASRP network.
| AP | AP 50 | AP s | AP m | AP l | |
|---|---|---|---|---|---|
| (a) DETR [ 36 ] | 29.2 | 60.2 | 11.6 | 38.0 | 47.4 |
| (b) (a)+SR+RDSP [ 6 ] | 30.1 | 61.4 | 13.6 | 38.3 | 47.4 |
| (c) (a)+ASSR [ 23 ] | 30.8 | 59.6 | 14.8 | 38.7 | 47.3 |
| (d) (c)+ASRP | 32.9 | 63.2 | 19.0 | 41.0 | 47.3 |
Table 1: Detection performance comparison. (a) DETR. (b) DETR is used in [ 6 ] , in which SR and region proposals with integer SR factors. (c) DETR is used with ASSR [ 23 ] . (d) DETR is used with ASSR and ASRP (ours).
Figure 4: Visual results. Green, blue, and red boxes indicate true positives, false negatives, and false positives, respectively. Five images at the bottom of each example are the heatmaps of (d).