Research map / Detection and segmentation / Open-set and few-shot detection
Shadow Generation with Decomposed Mask Prediction and Attentive Shadow Filling
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 DMASNet, a two-stage shadow generation network (decomposed box/shape mask prediction plus attention-based shadow filling), and a 280k-tuple rendered dataset RdSOBA to complement the small DESOBA dataset.
Problem
Generating plausible foreground shadows for inserted objects in composite images, which is limited by the small size of existing shadow datasets (DESOBA) and by artifacts such as holes and fragments in generated shadows that generalize poorly to real composites.
Approach
Decompose foreground shadow mask prediction into a bounding-box regression (relative to the object box) and a standardized-size shape prediction, then refine the mask; fill the shadow by scaling the composite so its shadow-region mean matches an attention-weighted mean of background shadow pixels; and pretrain on a large rendered dataset (RdSOBA) before finetuning on DESOBA.
Key results
- On RdSOBA+DESOBA training, DMASNet reaches S-RMSE 55.168 and S-PSNR 9.521 versus SGRNet's 59.855 and 8.921 on the DESOBA test set.
- DMASNet's d_hole drops to 0.212 (ground truth 1.076) with RdSOBA+DESOBA training, versus SGRNet's 5.276.
- In a 20-person user study on 100 real composite images, DMASNet with RdSOBA attains B-T score 2.039 versus SGRNet's -0.192.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DMASNet instead of SGRNet, trained on RdSOBA | Shadow generation · DESOBA test | d_frag: 30.686 -> 10.619 | Helped |
| Use DMASNet instead of SGRNet, trained on RdSOBA+DESOBA | Shadow generation · DESOBA test | S-RMSE: 59.855 -> 55.168 | Helped |
| Use DMASNet instead of SGRNet, trained on DESOBA | Shadow generation · DESOBA test | S-RMSE: 61.762 -> 65.847 | Hurt |
| Add RdSOBA training before fine-tuning on DESOBA | DMASNet · Real composite images | B-T score: 0.648 -> 2.039 | Helped |
| Add RdSOBA training before fine-tuning on DESOBA | SGRNet · Real composite images | B-T score: -0.104 -> -0.192 | Hurt |
| Train DMASNet on RdSOBA rather than DESOBA | DMASNet · DESOBA test | S-RMSE: 65.847 -> 63.693 | Helped |
| Train SGRNet on RdSOBA rather than DESOBA | SGRNet · DESOBA test | S-RMSE: 61.762 -> 66.603 | Hurt |
Surprising findings
- Without RdSOBA, DMASNet has worse RMSE/PSNR than SGRNet on DESOBA (e.g., RMSE 5.583 vs 4.754), yet is rated comparable or better visually.
- SGRNet barely improves when RdSOBA is added, while DMASNet improves substantially, indicating a transferability gap.
Limitations
- Pixel-wise RMSE/PSNR do not match visual quality; DMASNet trails SGRNet on these metrics.
- RdSOBA construction requires manual filtering of low-quality tuples.
- Real-image evaluation uses only 100 composites and a user study with 20 participants.
Experimental setup
Datasets: DESOBA RdSOBA real composite images (100, from prior work)
Benchmarks: DESOBA test set real composite images (100)
Metrics: RMSE S-RMSE PSNR S-PSNR BER S-BER d_hole d_frag B-T score
Baselines: Pix2Pix Pix2Pix-Res Mask-ShadowGAN ShadowGAN ARShadowGAN SGRNet
Builds on: Mask R-CNN Pix2Pix SGRNet CIoU loss
Closest papers by meaning
- Shadow Generation for Composite Image Using Diffusion Model (CVPR 2024 · similarity 0.89)
- Shadow Generation Using Diffusion Model with Geometry Prior (CVPR 2025 · similarity 0.88)
- MetaShadow: Object-Centered Shadow Detection, Removal, and Synthesis (CVPR 2025 · similarity 0.84)
- Polarization Guided Mask-Free Shadow Removal (AAAI 2025 · similarity 0.84)
- Detail-Preserving Latent Diffusion for Stable Shadow Removal (CVPR 2025 · similarity 0.84)
- Gaussian Shadow Casting for Neural Characters (CVPR 2024 · similarity 0.84)
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.