Research map / Detection and segmentation / CNN features

ContextShift: A Controlled Benchmark for Context Dependence in Object Detection

NeurIPS 2026 · Detection and segmentation › CNN features · approach: Region-based and query-based detectors · 0 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.

Introduces ContextShift, a controlled COCO-based benchmark that manipulates object-context relations (geometric, synthetic and NPMI-graded natural backgrounds) to show detectors suppress detections, and shows context-aware augmentation partially recovers performance.

Problem

Standard aggregate metrics like AP on uncontrolled distribution shifts conflate factors and obscure how object detectors degrade when object-context relationships change.

Approach

Treats context as an independent variable by manipulating object-context relations while preserving object appearance, with a continuous NPMI-based object-background compatibility axis and behavioral metrics (FN, FP, prediction volume) that reveal prediction suppression rather than confidence degradation.

Key results

• Natural background substitution (single-object setting) yields up to +227.1% FN increase (Faster R-CNN) and up to -44.2% prediction volume change (Def-DETR), with FP declining.
• Rotation yields the lowest implicit robustness (AP rAUC 0.57) and largest prediction-volume drop (-30.2% on average across models).
• Context-aware augmentation variants all outperform the dataset-only YOLO baseline on clean and manipulated data, improving AP@0.5 by 1.1-2.6% and reducing FN by 8.9-21.1%.

Surprising findings

• Performance along the NPMI compatibility axis is non-monotonic, peaking at intermediate compatibility and degrading at both extremes.
• Missed detections stem mainly from reduced candidate existence rather than reduced confidence, with false positives stable or declining.

Limitations

• Compatibility axis is based on statistical co-occurrence and does not capture visual cues like geometry, lighting, or perspective.
• Object-centric formulation restricts analysis to single objects; multi-object and relational settings are not covered.
• Evaluation limited to COCO 2017 and Places365 and to object detection.

Experimental setup

Datasets: COCO 2017 val COCO train2017 Places365

Benchmarks: ContextShift COCO 2017 val COCO-O

Metrics: AP@0.5 rAUC delta FN/img delta FP/img delta pred/img candidate existence rate

Baselines: Faster R-CNN YOLO26M Deformable DETR RF-DETR D-FINE COCO-O

Builds on: COCO-O COCO-C LaMa YOLO-World Places365

Closest papers by meaning

• Object Detection Benchmarks are Incomplete: The Role of Label Errors and Annotation Uncertainty (NeurIPS 2026 · similarity 0.88)
• ContextHOI: Spatial Context Learning for Human-Object Interaction Detection (AAAI 2025 · similarity 0.87)
• The Devil is in the Object Boundary: Towards Annotation-free Instance Segmentation using Foundation Models (ICLR 2024 · similarity 0.86)
• COUNTS: Benchmarking Object Detectors and Multimodal Large Language Models under Distribution Shifts (CVPR 2025 · similarity 0.86)
• Test-Time Adaptive Object Detection with Foundation Model (NeurIPS 2025 · similarity 0.86)
• Context Enhanced Transformer for Single Image Object Detection in Video Data (AAAI 2024 · similarity 0.85)