Arrow Detection With ConvNext Pretrained on DINOv3 Hits 96.6% F1
BrashBunyip · reddit · 2026-10-19
After a month of work, the author built an archery target impact-detection system: a registration network identifies the target, semantic rings, center, and scoring zones; a detector then produces a heatmap of impact candidates on the cropped image, run twice (mirrored) and fused in the decoder to find local peaks.
Data & models: 40,000 varied Blender-rendered scenes plus 8,000 hand-labeled real images (20 hours of labeling); both networks trained on 1M samples with heavy augmentation. ResNet (ImageNet-pretrained) performed well, pure ViT was worse, and ConvNext pretrained on DINOv3 was best.
Post-processing trick: homography removes perspective skew and flattens the target; then the predicted semantic-ring heatmap is compared against the ideal geometric rings to compute a radial offset correcting for warped (non-planar) paper—something homography alone can't fix.
Results: 96.6% F1 on the test set, 80.5% whole-image correctness (no missed arrows, no false detections).
More from Research
- EMNLP paper: diffusion LMs internally encode latent denoising timestep representations — gsarti_ · 2026-10-23
- KU Leuven and VIB open two Group Leader professorships in AI for Biology, start 2027 — jpillowtime · 2026-10-23
- Subspace Tuning curbs LLM catastrophic forgetting by freezing principal singular directions — scottniekum · 2026-10-23
- Perceptic launches PercepticOS, an AI workspace for 'scientific superintelligence' — nathanbenaich · 2026-10-23
- EPFL's Bosselut to talk long context as test-time learning at ELLIS seminar — andre_t_martins · 2026-10-23
- DRATS: Adaptive Task Sampling Boosts Worst-Task Performance in Multi-Task RL — Cohere · 2026-10-23