CV Brief · Saturday, 18 July 2026
CV Brief
Research & Papers
Zero-Shot Vision-Language Classification via Adaptive Prompt Reweighting
CARPRT optimizes prompt ensembling for VLMs by learning class-aware reweighting vectors without fine-tuning. This directly improves zero-shot classification accuracy on black-box models—critical for practitioners deploying CLIP and similar systems without training data.
Read more →LiDAR Terrain Features Boost Geospatial AI Explainability and Accuracy
Explainable geospatial AI leverages LiDAR-derived terrain intelligence to replace fixed land-use clutter heights with data-driven estimates. Practitioners working on satellite imagery, geospatial analysis, or remote sensing pipelines gain actionable terrain features and interpretability.
Read more →Multi-Domain Retrieval Control Certifies Per-Domain Contamination Bounds
C3R adds a drop-in control layer for retrieval systems that certifies per-domain contamination budgets without query-time labels, using conformal risk guarantees. Relevant for production retrieval pipelines mixing domains—ensures wrong-domain results don't silently degrade ranking.
Read more →Tools & Releases
Pipe defect detection: Train RF-DETR, automate QA workflows
Roboflow demonstrates automated tube and pipe inspection using RF-DETR to detect holes, cracks, and ruptures, with built-in pass/review/fail classification workflows. Directly applicable to manufacturing quality control pipelines where visual defect detection reduces manual inspection costs.
Read more →End-to-end porch lockbox: on-device detection + cloud validation
Complete build of a self-unlocking delivery box combining on-device iPhone detection with cloud Roboflow Workflows for courier verification and solenoid control via ESP32. Shows practical integration of edge inference, spatial reasoning, and hardware triggers—a real production system.
Read more →Retail shelf monitoring: RF-DETR product detection + LLM summaries
Detect visible beverage products with RF-DETR, compare counts against thresholds, and use Gemini 2.5 Pro for shelf inspection reports. Practical shelf-audit automation combining object detection with multimodal reasoning for retail operations.
Read more →Tutorials & Guides
GANs Still Outperform Diffusion Models in Specific Tasks
GANs remain competitive against diffusion models in particular use cases despite recent diffusion model popularity. The article explores where generative adversarial networks maintain advantages and remain production-viable for CV practitioners.
Read more →pHash deduplication for video crops: use Hamming distance ≤10 as your threshold. Too tight misses duplicates, too loose removes valid unique crops.