r28112072/upcount

Overview

Model overview: The model uses 384 x 384 training crops, a class-token-free ViT-B/16, intermediate blocks [2, 5, 8, 11], and a 64-channel stride-four refined feature map. FeatUp-style JBU restores image-aligned detail before the proposal and verification heads.

Training follows the CounTR-length two-stage FSC-147 schedule, followed by optional CARPK fine-tuning. Learning rates are linearly scaled as base_lr * effective_batch_size / 256, with AdamW and cosine decay.

Stage	Epochs	Batch	Base LR	Single-GPU LR	Warm-up	Weight decay
FSC-147 MAE pretraining	500	8	1.5e-4	4.6875e-6	10	0.05
FSC-147 counting fine-tuning	1000	26	2.0e-4	2.03125e-5	10	0.05
CARPK counting fine-tuning	1000	8	2.0e-4	6.25e-6	10	0.05

Full paper is available on arXiv, and code available on GitHub.

Citation

@article{wijaya2026spatially,
  title={Spatially-Aware Class-Agnostic Object Counting},
  author={Wijaya, Robert and Hossain, Md Tanvir and Kau, Amanda and Cheung, Ngai-Man},
  journal={arXiv preprint arXiv:2607.16826},
  year={2026}
}