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/

Global: Zendy Tools annual

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

Global: Zendy Tools monthly

Global: Zendy Plus monthly

Presents the access of premium content as premium feature

Premium Content

Global: Zendy Plus annual

Global: Zendy Free Plan

The introduction of DETR represents a new paradigm for object detection.However, its decoder conducts classification and box localization using sharedqueries and cross-attention layers, leading to suboptimal results. We observethat different regions of interest in the visual feature map are suitable forperforming query classification and box localization tasks, even for the sameobject. Salient regions provide vital information for classification, while theboundaries around them are more favorable for box regression. Unfortunately,such spatial misalignment between these two tasks greatly hinders DETR'straining. Therefore, in this work, we focus on decoupling localization andclassification tasks in DETR. To achieve this, we introduce a new design schemecalled spatially decoupled DETR (SD-DETR), which includes a task-aware querygeneration module and a disentangled feature learning process. We elaboratelydesign the task-aware query initialization process and divide thecross-attention block in the decoder to allow the task-aware queries to matchdifferent visual regions. Meanwhile, we also observe that the predictionmisalignment problem for high classification confidence and preciselocalization exists, so we propose an alignment loss to further guide thespatially decoupled DETR training. Through extensive experiments, wedemonstrate that our approach achieves a significant improvement in MSCOCOdatasets compared to previous work. For instance, we improve the performance ofConditional DETR by 4.5 AP. By spatially disentangling the two tasks, ourmethod overcomes the misalignment problem and greatly improves the performanceof DETR for object detection.

Decoupled DETR: Spatially Disentangling Localization and Classification for Improved End-to-End Object Detection