PaperlayerView exact revision on arXiv

Source-linked AI summary

Any-scale Object Detection using Arbitrary-scaled Images

Kazutoshi Akita, Norimichi Ukita

arXiv:2610.04346v1cs.CV

TL;DR

Discrete multi-scale representations may miss intermediate object scales, while preparing finer grids is impractical. The paper combines arbitrary-scale super-resolution with scale-specific region proposals to reduce false positives; the combination improves memory efficiency over.

  • Problem

    Discrete scale representations can miss intermediate-scale objects, while preparing finer-grained representations is impractical because of memory demands.

    Source: §1 · p. 1

  • Method

    The method combines arbitrary-scale super-resolution with scale-specific region proposals that select pixels for super-resolution at each scale, jointly trained with the detector.

    Source: §1 · p. 1

  • Results

    ASSR with ASRP improves memory efficiency over and is superior to the other methods on small-object APs, while large-object APl scores are almost the same.

    Source: §5 · p. 4, §4.3 · p. 4

  • Takeaways & Limitations

    Scale-specific proposals address false positives associated with applying ASSR across scales, supporting the paper's any-scale detection approach.

    Source: §1 · p. 1

    The method still leaves two false positives and one false negative unresolved in a qualitative example, and improving the heatmaps is identified as future work.

    Source: §4.4 · p. 4

Abstract

from arXiv · show

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.

1 Introduction

The paper targets scale gaps left by discrete image pyramids: finer scale grids are impractical, while intermediate-scale objects may be missed. It proposes arbitrary-scale super-resolution with scale-specific proposals and joint training across scales.

1 Introduction

  • Discrete pyramid scales can miss intermediate-scale objects, while preparing fine-grained representations is impractical because of memory demands.The paper gives a medium-large object scaled to ×1.5 as an example of a scale not represented by the described ×1/4, ×1, and ×4 choices.Source: §1 · p. 1
  • ASSR supports arbitrary-scale image scaling, but detections across scales can include false positives when local appearances resemble objects at other scales.Source: §1 · p. 1
  • Arbitrary-scale region proposals separate proposals by reconstructed scale, suppressing false positives while allowing continuous scales unlike the discretized proposals in.The paper illustrates scale-specific proposals for ×3.1, ×2.2, and ×1.6.Source: §1 · p. 1
  • The paper extends multi-scale detection to any-scale detection by optimizing ASSR for the object-detection task.Source: §1 · p. 1
  • Jointly training the region proposal and detector across scales is described as essential for generalizing from sparsely and discontinuously distributed training scales.Source: §1 · p. 1

2 Related Work

Related work covers arbitrary-scale super-resolution methods and multi-scale detection with image or feature pyramids. The paper distinguishes continuous scaling from integer-scale SR and places its approach among pyramid-based and SR-based detection methods.

2.1 Super-resolution

  • Many SR methods [7] [8] [9] [10] [11] [15] use a single integer scaling factor, whereas ASSR methods [17] [20] support any factor with one model.Source: §2.1 · p. 2
  • ASSR approaches predict upsampling filters, blend integer-scale SR images, or use LIIF-based implicit functions with variants that predict Fourier components, aggregate latent codes, or embed input positions.The passage names LTE and Thera as Fourier-component predictors, CiaoSR and CLIT as aggregators, and UltraSR and IPE as embedding approaches.Source: §2.1 · p. 2
  • recurrentLTE [29] trades SR quality against computational cost, allowing runtime control for applications such as intelligent-vehicle object detection.Source: §2.1 · p. 2

2.2 Multi-scale Object Detection

  • Detection work uses feature or image pyramids [1] [2] [3] [30] [4]; SR has also been applied to upsampled images, and Figure 2 contrasts multiple independent SR/proposal pairs with the proposed pair.Source: §2.2 · p. 2, Figure 2 · p. 2

3 Method

The method pairs arbitrary-scale super-resolution with scale-conditioned region proposals, then detects objects in the masked SR images. It trains proposal targets from annotated boxes and fine-tunes the detector for a common object scale.

3 Method

  • One ASSR–ASRP pair can run at any scale S ∈ R, while experiments sample {0.5, 1.0, 1.5, 2.0, 2.5} to cover test-image object variation at tractable computational cost.The paper says denser scale sampling remains straightforward within the framework.Source: §3 · p. 2

3.1 Arbitrary-scale Region Proposal Network

  • ASRP conditions its region-proposal heatmap on scale S, and the upscaled, thresholded heatmap masks the corresponding SR image before detection.Source: §3.1 · p. 3, §3 · p. 3
  • Training derives heatmap targets from annotated boxes at a 128-pixel target height; highlighting additional objects within an empirically chosen scale tolerance helps avoid single-object overfitting.The stated tolerance uses α = 0.7 and β = 1.3, and the heatmap values are blurred following.Source: §3.1 · p. 3

3.2 Object Detector

  • The pretrained detector is fine-tuned to detect objects at the common target height, and detections from all masked SR images are merged with nonmaximum suppression.Source: §3.2 · p. 3

4 Experiments

Experiments on CityScapes evaluate human-object detection and compare four detector configurations. Results report small-object performance, parameter counts, and qualitative detections across examples.

4.1 Dataset

  • The experiments use CityScapes, retaining person and rider classes and deriving bounding boxes from pixelwise instance labels.The train and test splits contain 2,965 and 492 images with human boxes, respectively.Source: §4.1 · p. 3

4.2 Implementation Details

  • Four DETR configurations are compared, including DETR with ASSR and ASRP; pretrained DETR and LTE are concatenated with ASRP and jointly trained on CityScapes.ASRP is trained from scratch in the end-to-end joint network.Source: §4.3 · p. 4, §4.2 · p. 4

4.3 Quantitative Results

  • Large-object scores (APl) are almost the same across methods, while our method is superior for small-object detection (APs).Source: §4.3 · p. 4

4.4 Qualitative Results

  • In the upper example, our method yields only a true positive where ASSR also produces a false positive; in the lower, errors remain.The lower example still contains two false positives and one false negative, and heatmaps for different scales overlap.Source: §4.4 · p. 4

5 Conclusion

The paper proposes any-scale object detection with ASSR and ASRP to suppress false positives. It reports improved memory efficiency over a prior discretized-scale method and identifies further evaluation and model exploration as future work.

  • ASSR supports any-scale object detection, while ASRP is proposed to suppress false positives that occur due to ASSR.Source: §5 · p. 4
  • The ASSR-ASRP combination improves memory efficiency compared with the previous method using discretized object scales.Source: §5 · p. 4
  • Future work includes evaluating generalization to other datasets, denser scale sampling, and different detectors and ASSR models.Source: §5 · p. 4
Loading 2610.04346v1…