The Map of CVPR 2026

Code behind "The Map of CVPR 2026" — a data-driven tour of all 5,010 accepted CVPR 2026 papers (4,069 main + 941 Findings). It scrapes the papers from CVF Open Access, embeds them with a sentence-transformer, clusters and projects them, and emits 7 interactive Plotly charts.

Charts

#	Chart	What it shows
01	Map	Every paper as a dot (UMAP of embeddings), colored by topic cluster; toggle main vs. Findings
02	Topics	The 24 clusters by paper count
03	Buzzwords	Share of papers mentioning each idea (diffusion, VLM, 3DGS, …)
04	Network	Co-occurrence of concept words in titles
05	Naming	Title linguistics — colon format, openers, reused acronyms
06	Trends	Year-over-year idea adoption across the 2024–2026 main tracks
07	Code	Open-source rate by topic cluster

Pipeline

pip install -r requirements.txt

# 1. Scrape the papers from CVF (main + Findings -> cvpr2026_papers.json/.csv)
python scrape_cvf.py

# 2. (optional) Prior years for the year-over-year trend chart
python scrape_cvf_years.py 2024 2025

# 3. Embed "title. abstract" for every paper (-> cvpr2026_embeddings.npz)
python build_embeddings.py

# 4. Build the 7 interactive charts (reuses a cached UMAP projection with --no-recompute)
python build_dashboard.py --out charts
#    add --images <dir> to also export static PNGs (needs kaleido)

cluster_labels.json holds the hand-written names for the 24 KMeans clusters (seed=42); delete it to fall back to automatic c-TF-IDF labels, or regenerate it if you change the embedding model or k.

Other tools

• search_cvpr.py — keyword / TF-IDF / semantic search over the scraped corpus
• analyze_cvpr.py — quick corpus stats (topics, co-occurrence, authors, clusters, dupes)

Notes

• Data is scraped fresh from CVF and is not committed (see .gitignore); run the pipeline above to regenerate it.
• Embeddings use all-MiniLM-L6-v2 by default. Proximity on the map is semantic text similarity, not citations or impact.

🤖 Generated with Claude Code