AI Paper Atlas

Research map / Label-efficient and robust learning / Recognition and long-tail

LoCo: Selective Local Competition for Discriminative Open-Vocabulary Multi-Label Recognition

Zhuoming Li, Bo Han, Xiaoyu Wang, Yuheng Jia

NeurIPS 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Partial and multi-label learning · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 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 LoCo, a selective local competition method for open-vocabulary multi-label recognition that uses local prediction as a discriminative branch.

Problem

Open-vocabulary multi-label recognition (identifying all queried concepts in an image) has focused on global image-text matching and text-side representations, leaving local prediction under-explored.

Approach

Revisits local prediction and shows it can act as a strong discriminative branch once spatial consistency and related properties are handled; introduces selective local competition to achieve this.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
replace CLIP with ClearCLIP dense features for local prediction (ablation)LoCo local branch · OV-MLR avg of 4F1 / mAP: 46.4 -> 52.5 F1; 55.0 -> 59.8 mAPImproved
use SC-CLIP instead of ClearCLIP (ablation)LoCo local branch · OV-MLR avg of 4mAP: 59.8 -> 62.4Improved
add Selective Softmax (ablation)LoCo (SC-CLIP, 14 crops) · OV-MLR avg of 4F1-score: 56.6 -> 58.5Improved
add Selective Softmax (ablation)LoCo (SC-CLIP, 14 crops) · OV-MLR avg of 4mAP: 65.5 -> 66.0Improved
add 14-crop multi-scale inference (ablation)LoCo (SC-CLIP + Sel) · OV-MLR avg of 4mAP: 62.9 -> 66.0Improved
use raw logits (no competition) instead of Selective Softmax (ablation)LoCo · OV-MLR avg of 4F1-score: 58.5 -> 53.1Hurt
use full softmax instead of Selective Softmax (ablation)LoCo · OV-MLR avg of 4F1-score: 58.5 -> 56.6Hurt
use random compatibility prior instead of LLM prior (ablation)LoCo · OV-MLR avg of 4F1-score: 58.5 -> 52.9Hurt
use WordNet prior instead of LLM prior (ablation)LoCo · OV-MLR avg of 4F1-score: 58.5 -> 57.0Hurt
use LLM consensus prior instead of single GPT-5.5 prior (ablation)LoCo · OV-MLR avg of 4F1-score: 58.5 -> 58.6No clear effect
add multi-scale inference to TaI-DPT with SC-CLIP (ablation)TaI-DPT local prediction · OV-MLR avg of 4F1-score: 55.8 -> 55.7No clear effect
apply LoCo with CLIP backbone instead of SC-CLIP (ablation)LoCo · OV-MLR avg of 4mAP: 66.0 -> 55.3Hurt

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Surprising findings

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy