Benchmark RadarAI BENCHMARK PROFILE
ContextShift
ContextShift is a controlled benchmark that manipulates object-context relationships in COCO images to isolate context as an independent variable for object detection evaluation.
- Released
- 2026-06-08
- Readiness
- Paper only
- Primary field
- General AI
Why it matters
It reveals that standard aggregate metrics like AP can mask substantial recall loss and changes in prediction dynamics under context variation, aiding in understanding detector robustness.
Motivation
Modern object detectors achieve strong performance on standard benchmarks, yet their robustness to contextual variation remains insufficiently understood.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.