Research map / Image generation / Autoregressive and latent generation
Spatial-Semantic Collaborative Cropping for User Generated Content
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes S²CNet, a spatial-semantic adaptive attention graph for aspect-ratio-constrained image cropping of user generated content, plus the UGCrop5K benchmark.
Problem
Image cropping for user generated content (UGC) with multiple objects and cluttered backgrounds, requiring both aesthetics and content integrity under fixed aspect ratios.
Approach
Builds a fully connected graph over detected object RoIs and the crop candidate, with semantic (appearance) and spatial (center-distance) edges, and uses a graph-aware attention transformer to propagate information to the crop node for aesthetic score regression.
Key results
- On UGCrop5K, S²CNet reaches SRCC 0.502, ACC@5 60.8 and ACC@10 72.1, above the prior best TransView (57.9 / 69.4 on ACC).
- On GAIC v2, S²CNet achieves ACC@5 64.0 and ACC@10 82.7, with 3.92M parameters and 162.8 FPS on a single RTX 2080Ti.
- In a user study with 15 volunteers, S²CNet was preferred in 47.7% (3:4), 44.0% (4:3) and 52.0% (16:9) of cases, the highest share among compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use S²CNet instead of the strongest listed baseline, TransView | image cropping · UGCrop5K | averaged ACC5: 57.9 -> 60.8 | Helped |
| Use S²CNet instead of CGS | image cropping · GAICv1 | averaged ACC5: 59.7 -> 61.0 | Helped |
| Use S²CNet instead of SFRC | image cropping · GAICv2 | averaged ACC10: 82.6 -> 82.7 | Helped |
| Add FAG, semantic edges, and spatial edges to the Transformer baseline (ablation) | S²CNet · UGCrop5K | averaged ACC5: 54.3 -> 60.8 | Helped |
| Add FAG, semantic edges, and spatial edges to the Transformer baseline (ablation) | S²CNet · GAICv1 | averaged ACC5: 49.7 -> 61.0 | Helped |
| Use learned spatial embeddings (DisEmb) rather than DisDrop (ablation) | S²CNet · UGCrop5K | averaged ACC5: 60.4 -> 60.8 | Helped |
| Use 8 rather than 10 object proposals (ablation) | S²CNet · UGCrop5K | averaged ACC5: 60.8 -> 59.9 | Hurt |
| Use 12 rather than 10 object proposals (ablation) | S²CNet · UGCrop5K | averaged ACC5: 60.8 -> 60.6 | Hurt |
| Use 15 rather than 10 object proposals (ablation) | S²CNet · UGCrop5K | averaged ACC5: 60.8 -> 60.4 | Hurt |
| Replace the adaptive attention graph with GCN (ablation) | S²CNet · UGCrop5K | averaged ACC5: 60.8 -> 56.2 | Hurt |
| Replace the adaptive attention graph with GAT (ablation) | S²CNet · UGCrop5K | averaged ACC5: 60.8 -> 59.4 | Hurt |
| Replace the adaptive attention graph with GCN (ablation) | S²CNet · GAICv1 | averaged ACC5: 61.0 -> 57.3 | Hurt |
Surprising findings
- Replacing the spatial distance drop (DisDrop) with a learned distance embedding (DisEmb) was consistently better, suggesting distant nodes still carry useful relations.
- Adding the spatial edge alone improved performance more than adding the semantic edge alone on the ablation.
Limitations
- UGCrop5K experiments used an early version of the dataset annotated by partial annotators; the full dataset is promised later.
- Only 90 predefined anchor crops per image and a limited number of annotators (20) bound the annotation coverage.
- Baseline results for ASM-Net and TransView were re-implemented by the authors because code was unavailable.
Experimental setup
Datasets: UGCrop5K GAIC v1 GAIC v2 Visual Genome KoNViD-1k LIVE-VQC YouTube-UGC Bilibili
Benchmarks: UGCrop5K GAIC v1 GAIC v2
Metrics: SRCC ACC@5 ACC@10 FPS user preference percentage
Baselines: VFN A2-RL VEN VPN GAICv1 GAICv2 ASM-Net CGS TransView HCIC SFRC Mars
Builds on: Transformer GCN GAT Faster R-CNN RoIAlign
Closest papers by meaning
- Image Cropping With Spatial-Aware Feature and Rank Consistency (CVPR 2023 · similarity 0.89)
- Reliable and Efficient Image Cropping: A Grid Anchor Based Approach (CVPR 2019 · similarity 0.88)
- Learning Subject-Aware Cropping by Outpainting Professional Photos (AAAI 2024 · similarity 0.86)
- Cropper: Vision-Language Model for Image Cropping through In-Context Learning (CVPR 2025 · similarity 0.85)
- Venus: Benchmarking and Empowering Multimodal Large Language Models for Aesthetic Guidance and Cropping (CVPR 2026 · similarity 0.85)
- SSMG: Spatial-Semantic Map Guided Diffusion Model for Free-Form Layout-to-Image Generation (AAAI 2024 · similarity 0.82)
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.