CV Brief · Monday, 3 August 2026
CV Brief
Research & Papers
Metal artifact reduction in CT via structure-conditioned flow matching
SCMA tackles beam hardening and metal scatter artifacts in CT imaging using flow matching conditioned on anatomical structure. This directly addresses a clinical imaging bottleneck where metal implants degrade diagnostic quality, offering practitioners a generative approach that preserves structural fidelity better than optimization or regression baselines.
Read more →Frequency-aware flow matching improves image generation coherence and detail
WaiT decomposes image generation into frequency bands using wavelet transforms, treating high and low frequencies differently in flow matching. For practitioners scaling generative models to high resolution, this offers a tractable way to improve both global structure and texture fidelity without proportional compute overhead.
Read more →Synthetic data fails to close generalization gap in plant phenotyping detection
This empirical study quantifies how genotype-by-environment shifts degrade object detection in cowpea breeding across locations and seasons, showing synthetic data alone cannot close the generalization gap. For agricultural CV practitioners, it signals that domain shift in real-world phenotyping requires mixed real/synthetic strategies, not synthetic-only solutions.
Read more →Tutorials & Guides
Plant disease detection from single photo: real failures, practical lessons
Engineer built CV system to identify plant diseases from images and documented both successes and failure modes. Practical walkthrough of dataset challenges, model selection, and production edge cases relevant to agricultural CV applications.
Read more →Jumping jack counter with pose estimation: production debugging guide
Real-world case study building rep-counting system with MediaPipe pose estimation on mobile video. Breaks down actual failures encountered—occlusion, angle variance, temporal noise—and solutions tested on production phone footage.
Read more →Industry & Deployments
Pokémon GO geospatial data pipeline: crowdsourced imagery at scale
Analysis of how Pokémon GO mobilized millions as unpaid data collectors for geospatial CV training. Examines annotation quality, coverage bias, and implicit incentive structures in crowd-sourced street-level imagery pipelines.
Read more →When extracting crops from CCTV at scale, always use frame seeking (cv2.CAP_PROP_POS_FRAMES) instead of sequential reads. On a 2-hour video at 1FPS you'll go from hours to minutes.