File size: 1,292 Bytes
bde2b27
 
 
dee6a1f
bde2b27
c772f74
 
b27fd73
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e655232
c772f74
 
 
 
 
 
 
 
 
 
bde2b27
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
---
tags:
- object-counting
pipeline_tag: object-detection
---
## 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](https://arxiv.org/abs/2607.16826), and code available on [GitHub](https://github.com/r28112072-rgb/upcount).

## 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}
}
```