Frozen Vision Transformers for Dense Prediction on Small Datasets: A Case Study in Arrow Localization
Maxwell Shepherd · 2026
Read Frozen Vision Transformers for Dense Prediction on Small Datasets: A Case Study in Arrow Localization by Maxwell Shepherd (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-16758.txt.
Description
Shepherd presents a system detecting, localizing and scoring arrow punctures on archery target faces from only 48 photographs. The pipeline rectifies distorted photos into a standard system, then pairs a frozen DINOv3 transformer with guided upsampling. Only about 4 million of 308 million parameters are trainable, yet F1 reaches about 89 percent with localization error of one point four one millimetres. An ablation shows the CenterNet offset head adds negligible detection gain while degrading localization. Per-image scores return with median error of one point eight percent, and group centroids land within 4 millimetres.
Catalog record
| CREATOR | Maxwell Shepherd |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 27703 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.16758