Research map / Image generation / Text-to-image diffusion
AFreeCA: Annotation-Free Counting for All
Adriano D'Alessandro, Ali Mahdavi-Amiri, Ghassan Hamarneh
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add density-classifier-guided partitioning (DCGP 1x1 to 3x3) at inference (ablation) | AFreeCA counting network · ShanghaiTech B | MAE: 42.1 -> 35.0 | Improved |
| partition every image 3x3 (fixed) instead of density-guided partitioning (ablation) | AFreeCA counting network · UCF-QNRF | MAE: 392.5 (1x1) -> 425.6 (3x3 fixed) | Hurt |
| use regional density guidance instead of an image-level density decision for partitioning (ablation) | AFreeCA · UCF-QNRF | MAE: 309.5 -> 283.1 | Improved |
| add sorting pre-training with both l_sort^y and l_sort^z instead of ImageNet pre-training (ablation) | AFreeCA · ShanghaiTech B | MAE: 68.1 -> 35.0 | Improved |
| add feature-space ranking loss l_sort^z to the sorting loss (ablation) | AFreeCA sorting pre-training · ShanghaiTech A | MAE: 188.4 -> 152.7 | Improved |
| finetune only the last linear layer instead of the entire sorting network (ablation) | AFreeCA counting network · ShanghaiTech B | MAE: 43.9 -> 35.0 | Improved |
| sample patches from the original high-resolution image rather than a low-resolution one (ablation) | AFreeCA partitioning · UCF-QNRF | MAE: 317.6 -> 283.1 | Improved |
| filter outliers in the synthetic counting data with prototype similarity (ablation) | AFreeCA counting data · ShanghaiTech B | MAE: 48.2 -> 35.0 | Improved |
| use AFreeCA instead of CrowdCLIP on ShanghaiTech A | unsupervised crowd counting · ShanghaiTech A | MAE: 146.1 (CrowdCLIP) vs 152.7 | Hurt |
| fixed 2x2 partitioning of all images on ShanghaiTech B (ablation) | AFreeCA · ShanghaiTech B | MAE: 42.1 -> 51.6 | Hurt |
Closest papers by meaning
- Learning to Count without Annotations (CVPR 2024 · similarity 0.91)
- A Fixed-Point Approach to Unified Prompt-Based Counting (AAAI 2024 · similarity 0.88)
- OmniCount: Multi-label Object Counting with Semantic-Geometric Priors (AAAI 2025 · similarity 0.88)
- Make It Count: Text-to-Image Generation with an Accurate Number of Objects (CVPR 2025 · similarity 0.87)
- CountGD: Multi-Modal Open-World Counting (NeurIPS 2024 · similarity 0.87)
- T2ICount: Enhancing Cross-modal Understanding for Zero-Shot Counting (CVPR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
