Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →brycewang-stanford/cvpr-topic-selection
Use when deciding whether a project belongs at CVPR or should route elsewhere, covering what counts as a vision contribution at the field's flagship, fit tests for methods, datasets, and application papers, realistic odds at 25% acceptance and 16k submissions, and routing to ICCV, ECCV, WACV, 3DV, NeurIPS, or a journal.
Passed 20 of 23 test cases with the skill, and 14 of 23 without it.
and 5 more in the full report
6 fixed14 passed either way3 still failing0 broke
Tokens, turns and time describe resource use, not a dollar price.
One of 6 test cases the skill fixed. Same model, same prompt, with and without the skill.
We created a hardware-independent sparse tensor kernel that accelerates 3D point cloud convolutions, shifting the accuracy-per-FLOP frontier by 30% across multiple GPU architectures with full CRF reporting. Evaluate venue fit and output the structured assessment.
| Checked | Without skill | With skill |
|---|---|---|
| Evaluates hardware-independent throughput shifts as valid fit evidence. | Passed | Passed |
| Outputs a binary yes or no status for process tax readiness. | Missed | Passed |
The instructions your model reads when the skill loads.
CVPR is the largest venue in computer vision and one of the largest in all of science — 16,092 reviewed submissions and 4,090 acceptances in 2026. Size cuts both ways: almost any vision-adjacent topic has a reviewer pool there, and almost any weakness has a reviewer who has seen it a hundred times. This skill decides whether to feed the machine before other skills decide how.
Strip the engineering and ask: is the contribution a claim about visual data or visual computation? CVPR's 2026 program clustered exactly there — the largest areas were image/video synthesis and generation; vision+language and reasoning; multimodal learning; 3D from multi-view and sensors; and medical/biological vision (official program announcement). Contributions where vision is merely the demo domain — a generic optimizer tested on ImageNet, an ML theory result with a CIFAR table — historically route better to NeurIPS/ICML, where the reviewer pool evaluates the actual claim.
| You have… | CVPR-shaped if… | Warning sign |
|---|---|---|
| A method/architecture | It solves a visual mechanism (geometry, temporal, pixels-to-structure), with benchmark wins + ablations | Gain vanishes under matched backbones |
| A dataset/benchmark | It unlocks a task the field cannot currently study, with baselines and analysis | "Bigger than the last one" is the whole pitch (and release is due at camera-ready) |
| A systems/efficiency result | Accuracy-per-FLOP frontier moves; CRF-style reporting is your friend | Speedup only on your hardware story |
| A vision-language model result | The visual grounding is the contribution | It's an LLM paper wearing an image encoder |
| An application (medical, agriculture, driving) | A general vision insight travels beyond the application | Domain novelty only → domain venue or WACV |
| Theory about vision | Predicts something checkable in experiments | Pure theory → NeurIPS/ICML/SIGGRAPH-adjacent |
cvpr-related-work first.)test cases passed, vs 14 / 23 without
Published skills whose descriptions closely match this one, with their headline benchmark lift.
More skills from the same source and nearby in the registry, with their headline benchmark lift.