AI Paper Atlas

Research map / Detection and segmentation / CNN features

CoDeRNet: Selective Cross-Task Routing under Heterogeneous Supervision for Change Detection and Captioning

Eunki Cho, Hyeon Bae Kim, Seong Tae Kim

NeurIPS 2026 · Detection and segmentation › CNN features · approach: Domain-specific segmentation and detection · 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 CoDeRNet, a unified change detection and captioning model for bi-temporal remote sensing images that uses selective cross-task routing to handle heterogeneous supervision between localization and description.

Problem

Joint change detection and captioning (CDC) on bi-temporal remote sensing images, where localization and language description receive heterogeneous supervision and task-specific communication between them is under-explored.

Approach

Selective cross-task routing that controls how detection and captioning branches exchange information under heterogeneous supervision, rather than relying only on shared representations or jointly optimized architectures.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add CoDeR routing to joint detection and captioning training (ablation)CoDeRNet joint model · LEVIR-MCICIDEr-D: 136.18 -> 141.79Improved
add CoDeR routing to joint training (ablation)CoDeRNet joint model · LEVIR-MCImIoU: 87.67 -> 87.79Improved
joint training without CoDeR versus captioning-only training (ablation)captioning branch · LEVIR-MCICIDEr-D: 137.41 -> 136.18Hurt
joint training without CoDeR versus detection-only training (ablation)detection branch · LEVIR-MCImIoU: 87.03 -> 87.67Improved
use bidirectional routing instead of cap to seg only (ablation)CoDeR · LEVIR-MCICIDEr-D: 138.62 -> 141.79Improved
use bidirectional routing instead of seg to cap only (ablation)CoDeR · LEVIR-MCICIDEr-D: 141.95 -> 141.79No clear effect
use decomposed mask (1-C_t)*C_s instead of sender reliability only (ablation)CoDeR · LEVIR-MCICIDEr-D: 139.57 -> 141.79Improved
use decomposed mask instead of receiver need only (ablation)CoDeR · LEVIR-MCICIDEr-D: 141.02 -> 141.79Improved
use decomposed mask instead of a single gate (ablation)CoDeR · LEVIR-MCICIDEr-D: 138.19 -> 141.79Improved
use CoDeR instead of co-attention interaction (ablation)CoDeRNet · LEVIR-MCICIDEr-D: 139.33 -> 141.79Improved
use CoDeR instead of co-attention for detection (ablation)CoDeRNet · LEVIR-MCImIoU: 87.84 (Co-Attn) vs 87.79 (CoDeR)No clear effect
use CoDeR instead of gated modulation (ablation)CoDeRNet · LEVIR-MCICIDEr-D: 140.86 -> 141.79Improved

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