Research map / Detection and segmentation / CNN features

Size-invariance Matters: Rethinking Metrics and Losses for Imbalanced Multi-object Salient Object Detection

ICML 2024 · Detection and segmentation › CNN features · approach: RGB-D salient object detection · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes size-invariant SOD metrics (SI-MAE, SI-F, SI-AUC) that evaluate each salient object separately and a size-invariant loss wrapper (SI-SOD) applied to five SOD backbones, with a generalization bound.

Problem

Standard SOD metrics (MAE, F-score, AUC) are size-sensitive when multiple salient objects of diverse sizes co-exist, so small objects are overlooked in evaluation and training.

Approach

Rewrite separable and composite metrics as size-weighted sums over parts and replace the size-dependent weight with a constant, so each connected salient object (plus a background frame) counts equally; use the same size-invariant decomposition as a training loss, and derive a Rademacher-based generalization bound.

Key results

• On MSOD, adding the size-invariant loss improves EDN's SI-MAE from 0.0788 to 0.0724 and SI-F max from 0.8321 to 0.8619.
• On MSOD, averaged over backbones, the method outperforms baselines by about 0.012 on SI-MAE, 0.038 on SI-AUC, 0.070 on SI-F mean, 0.065 on SI-F max and 0.038 on E-measure.
• On DUTS-TE the method achieves similar MAE for single-object-dominant data, with small average gains (about 0.001 SI-MAE, 0.024 SI-F mean).

Measured effects reported in the paper

Surprising findings

• Original MAE can be equal for two predictions where the one that misses an entire small object is not penalized more, motivating size-invariant evaluation.
• The proposed loss also improves the original MAE on the multi-object MSOD dataset for all backbones.

Limitations

• Instance partition relies on connected components of the ground-truth mask as a proxy for objects, since no instance labels exist.
• Gains on single-object DUTS-TE are small and the method is competitive rather than clearly superior on original MAE and F-score.

Experimental setup

Datasets: DUTS-TR DUTS ECSSD DUT-OMRON HKU-IS MSOD PASCAL-S SOD XPIE

Benchmarks: MSOD DUTS-TE ECSSD DUT-OMRON HKU-IS PASCAL-S SOD XPIE

Metrics: MAE SI-MAE AUC SI-AUC F-measure (mean) SI-F (mean) F-measure (max) SI-F (max) E-measure

Baselines: PoolNet LDF ICON GateNet EDN

Builds on: DiceLoss IOU loss BCE loss vector contraction inequality

Closest papers by meaning

• Revisiting Salient Object Detection: Simultaneous Detection, Ranking, and Subitizing of Multiple Salient Objects (CVPR 2018 · similarity 0.88)
• A Causal Debiasing Framework for Unsupervised Salient Object Detection (AAAI 2022 · similarity 0.88)
• Modeling the Distributional Uncertainty for Salient Object Detection Models (CVPR 2023 · similarity 0.87)
• Alignment-Free RGB-T Salient Object Detection: A Large-Scale Dataset and Progressive Correlation Network (AAAI 2025 · similarity 0.85)
• ContextShift: A Controlled Benchmark for Context Dependence in Object Detection (NeurIPS 2026 · similarity 0.85)
• DiMSOD: A Diffusion-Based Framework for Multi-Modal Salient Object Detection (AAAI 2025 · similarity 0.85)