Research map / Point clouds / Point cloud classification
Enabling Region-Specific Control via Lassos in Point-Based Colorization
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.
Introduces a lasso tool and a localized cross-attention colorization model that masks color-hint attention to user-drawn regions, reducing color collapse in point-based interactive colorization.
Problem
Point-based interactive colorization suffers from "color collapse" when different color hints are given to semantically similar regions, because points alone cannot define each color's spatial extent.
Approach
Adds a loosely drawn lasso per color hint and uses it to build a localization attention mask that restricts cross-attention between grayscale image tokens and hint tokens, trained with simulated lassos (random rectangles around hints).
Key results
- A single lasso is reported to be as effective as about 4.18 color points in a user study (15.12 points replaced by 3.64 lassos).
- Users reach the same quality about 30% faster with lassos than with points alone (29.50 seconds faster).
- On the synthetic color-collapse dataset, the ablation shows PSNR@10 of 29.987 for the full model versus 26.728 without the localization mask.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the localization attention mask during training (ablation) | Lasso colorization model · ImageNet ctest | PSNR@1: 25.805 -> 26.722 | Helped |
| Add the localization attention mask during training (ablation) | Lasso colorization model · ImageNet ctest | PSNR@10: 26.728 -> 29.987 | Helped |
| Add the localization attention mask during training (ablation) | Lasso colorization model · ImageNet ctest | PSNR@100: 30.190 -> 33.349 | Helped |
| Add lasso interactions | Point-based colorization model · user study | time to target PSNR: 29.50 seconds faster | Helped |
| Replace some point interactions with lasso interactions | Point-based colorization model · user study | points replaced per lasso: 4.18 | Helped |
| Vary predefined lasso size across r:0.25, r:0.5, r:1, r:4, r:16, and r:64 (ablation) | Lasso colorization model · ImageNet ctest | PSNR: PSNR@1: 26.716, 26.720, 26.722, 26.723, 26.317, 26.003; PSNR@10: 29.965, 29.975, 29.987, 29.961, 28.802, 28.477; PSNR@10 | Mixed |
| Use lasso interactions | Point-based colorization model · user study | participant-reported benefit: 85.7\% | Helped |
Surprising findings
- Without lassos, the model performs comparably to iColoriT on the standard benchmark, so the lasso gain is specific to collapse-prone cases.
Limitations
- Lasso precision is limited by the attention-map resolution (H/P x W/P), so fine details need point refinement.
- Training-free lasso application to a model trained without the mask performs worse than training with it.
Experimental setup
Datasets: ImageNet 2012 ImageNet ctest Oxford 102 flowers CUB-200 Unsplash samples (98 collected)
Benchmarks: ImageNet ctest Oxford 102 flowers CUB-200 synthetic 2x2 color collapse dataset Unsplash collected dataset
Metrics: PSNR LPIPS HPR user study PSNR over time interactions required
Baselines: iColoriT UniColor Zhang et al. point-based colorization Su et al. (adaptive colorization) Kim et al.
Builds on: iColoriT Transformer decoder Huber loss pixel shuffle
Closest papers by meaning
- Instance-Aware Image Colorization (CVPR 2020 · similarity 0.86)
- Learning a Deep Convolutional Network for Colorization in Monochrome-Color Dual-Lens System (AAAI 2019 · similarity 0.85)
- Colorization Transformer (ICLR 2021 · similarity 0.85)
- Automatic Controllable Colorization via Imagination (CVPR 2024 · similarity 0.85)
- MangaNinja: Line Art Colorization with Precise Reference Following (CVPR 2025 · similarity 0.84)
- Color-Oriented Redundancy Reduction in Dataset Distillation (NeurIPS 2024 · similarity 0.83)
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.