English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Zendy Plus

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Zendy Tools

Zendy Open

Human-Object Interaction (HOI) detection is a core task for high-level imageunderstanding. Recently, Detection Transformer (DETR)-based HOI detectors havebecome popular due to their superior performance and efficient structure.However, these approaches typically adopt fixed HOI queries for all testingimages, which is vulnerable to the location change of objects in one specificimage. Accordingly, in this paper, we propose to enhance DETR's robustness bymining hard-positive queries, which are forced to make correct predictionsusing partial visual cues. First, we explicitly compose hard-positive queriesaccording to the ground-truth (GT) position of labeled human-object pairs foreach training image. Specifically, we shift the GT bounding boxes of eachlabeled human-object pair so that the shifted boxes cover only a certainportion of the GT ones. We encode the coordinates of the shifted boxes for eachlabeled human-object pair into an HOI query. Second, we implicitly constructanother set of hard-positive queries by masking the top scores incross-attention maps of the decoder layers. The masked attention maps then onlycover partial important cues for HOI predictions. Finally, an alternatestrategy is proposed that efficiently combines both types of hard queries. Ineach iteration, both DETR's learnable queries and one selected type ofhard-positive queries are adopted for loss computation. Experimental resultsshow that our proposed approach can be widely applied to existing DETR-basedHOI detectors. Moreover, we consistently achieve state-of-the-art performanceon three benchmarks: HICO-DET, V-COCO, and HOI-A. Code is available athttps://github.com/MuchHair/HQM.

Towards Hard-Positive Query Mining for DETR-based Human-Object Interaction Detection