CV Brief · Friday, 18 September 2026
CV Brief
Research & Papers
Selective prediction with uncertainty for cytology screening systems
Framework for deferring uncertain predictions in deep learning-based cervical cytology screening rather than forcing classification on every slide. Evaluates systems by ranking errors to bottom of confidence scores, matching real clinical deployment where human experts handle uncertain cases. Critical for medical imaging systems where confident decisions matter more than coverage.
Read more →Zero-shot object counting via density and point consistency models
DualCount addresses zero-shot counting by combining density regression with instance prediction, fixing spatial ambiguity and background leakage problems in weakly-supervised density approaches. Handles complex scenes without category-specific training—directly applicable to variable object counting in production pipelines.
Read more →Shadow harmonization for face compositing without identity loss
Multiplicative, albedo-preserving relighting pipeline for photometric plausibility in face compositing without re-synthesis that risks altering identity or skin tone. Solves practical problem in face-swap pipelines where donor and host carry mismatched illumination.
Read more →Tools & Releases
GPT-6 Astra polygons vs SAM 3: segmentation benchmark
GPT-6 Astra generates polygon outlines from text descriptions—not pixel masks. Roboflow benchmarked it against SAM 3 and shows how to combine both approaches for complementary strengths in segmentation pipelines.
Read more →Tutorials & Guides
Multimodal AI: Connecting Vision and Language Models
Explores how NLP and Computer Vision integrate to move beyond pixel processing toward semantic understanding. Directly relevant for practitioners building multimodal systems that combine OCR, object detection, and language understanding in production pipelines.
Read more →State Recovery for Long-Running Vision Automation Systems
Addresses failure modes in sustained automation tasks where single-frame failures cascade into system-wide context loss. Critical insight for teams deploying CV pipelines in game automation, robotics, or continuous monitoring where state management determines reliability.
Read more →Industry & Deployments
Urban-Scale Computer Vision: Morphological Analysis from Street Imagery
Demonstrates computational approaches to analyzing cities through visual data, from street-level imagery to urban morphology. Useful for practitioners working on geospatial CV projects, city planning systems, or large-scale visual annotation pipelines.
Read more →Auto-labeling confidence threshold: don't use 0.5. For quality training data, start at 0.7 and manually review the 0.5–0.7 band. The borderline cases are where your model learns.