Radar
Papers and tools are cheap. Runnable systems are not. This bridge tracks applied CV/ML ideas by can you run it, how hard, and what blocks shipping - with links into the Lab when we already teach it.
21 entries
paper · cv ·
Any-view geometry from one or many images - depth, pose, and 3DGS-ready space recovery.
Why care: Current frontier of the Depth Anything line: multi-view consistent geometry without a bespoke architecture zoo.
paper · cv ·
Temporally consistent depth for long videos, built on Depth Anything V2.
Why care: The bridge from single-frame depth demos to usable video / robotics / editing pipelines.
tool · cv ·
Current Ultralytics YOLO stack for detect / segment / pose / classify with a one-liner train and export path.
Why care: The default answer when someone asks “can we detect X in a camera stream?” in 2024+.
paper · cv ·
Promptable segmentation for images and video with temporal mask propagation.
Why care: The practical SAM successor when the job is video editing, tracking, or interactive labeling over time.
paper · cv ·
Stronger synthetic-teacher monocular depth with efficient small/base/large variants.
Why care: Still the workhorse monocular depth model most demos and ControlNet pipelines wire up first.
paper · cv ·
First Depth Anything release - strong monocular depth as a drop-in perception prior.
Why care: Historical baseline of the series - prefer V2 / DA3 for new work.
paper · cv ·
Promptable segmentation foundation model for masks from points/boxes.
Why care: Great for labeling acceleration and interactive tools - heavy for naive real-time video. Prefer SAM 2 for video.
paper · cv ·
Self-supervised vision backbone useful for retrieval and few-shot heads.
Why care: Strong frozen features when you lack labels - not always better than a tuned YOLO for detection.
paper · cv ·
Add spatial controls (edges, pose, depth) to diffusion image generation.
Why care: Turns “pretty random images” into directed product imagery pipelines.
tool · nlp ·
Open speech recognition that is good enough for many meeting/notes workflows.
Why care: Fast path to audio → text products; hallucinations need post-filters.
technique · cv ·
Detect text blocks, tables, and figures before OCR so structure survives.
Why care: The difference between “searchable PDF” and a usable knowledge base.
technique · nlp ·
Parameter-efficient fine-tuning so you can specialize a model without full retrain cost.
Why care: The practical path for domain tone/format - easy to overfit a tiny dataset.
paper · cv ·
Contrastive image–text model that classifies with natural language labels.
Why care: Useful when you don’t have a labeled dataset yet - weak when classes are visually subtle.
technique · nlp ·
Chunk docs, embed, retrieve top-k, stuff into an LLM prompt with citations.
Why care: The default pattern for private knowledge - most demos skip evaluation.
tool · mlops ·
Annotation UI + export loop so CV/NLP datasets can improve weekly.
Why care: Models stall when labels stall - tooling matters as much as architecture.
tool · cv ·
On-device landmark tracking for pose and hands without a custom detector stack.
Why care: Fast path for interaction demos and rehab/fitness prototypes.
tool · mlops ·
Thin HTTP wrapper around a model with health checks, timeouts, and versioned artifacts.
Why care: Notebook → API is where most “demos” die; this is the bridge.
technique · cv ·
Document OCR that actually works when you fix deskew, contrast, and layout first.
Why care: Most “OCR fails” are preprocessing fails, not model fails.
technique · cv ·
Single-shot detectors that made real-time object detection practical on commodity hardware.
Why care: Historical root of the YOLO line - for what people run today, see YOLO11 below.
technique · cv ·
Thresholding, contours, morphology, and geometric transforms - the boring stack that still ships.
Why care: Beats a neural net when lighting is controlled and you need deterministic industrial behavior.
technique · mlops ·
ColumnTransformer + Pipeline + model card metrics before you reach for deep learning.
Why care: If tabular baselines win, ship them - save GPUs for when they earn their keep.