NTIRE 2025 Image Shadow Removal Challenge Report

Florin-Alexandru Vasluianu, Tim Seizinger, Zhuyun Zhou, Cailian Chen, Zongwei Wu, Radu TimofteMingjia LiJin HuHainuo WangHengxing LiuJiarui WangQiming HuXiaojie GuoXin LuJiarong YangYuanfei BaoAnya HuZihao FanKunyu WangJie XiaoXi WangXueyang FuZheng-Jun ZhaYu-Fan LinChia-Ming LeeChih-Chung HsuXingbo · 2025

Read NTIRE 2025 Image Shadow Removal Challenge Report by Florin-Alexandru Vasluianu, Tim Seizinger, Zhuyun Zhou, Cailian Chen, Zongwei Wu, Radu TimofteMingjia LiJin HuHainuo WangHengxing LiuJiarui WangQiming HuXiaojie GuoXin LuJiarong YangYuanfei BaoAnya HuZihao FanKunyu WangJie XiaoXi WangXueyang FuZheng-Jun ZhaYu-Fan LinChia-Ming LeeChih-Chung HsuXingbo (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-15524.txt.

Introduction

Vasluianu leads a very large team reporting the NTIRE 2025 Image Shadow Removal Challenge, in a 2025 paper. Shadow removal is a step in cleaning up photos and computer vision inputs. The challenge drew hundreds of registered participants.

Seventeen teams submitted full solutions for final scoring. The contest had two tracks, one judging how faithful the rebuilt image is, the other judging visual quality through a user study. Test images came from the WSRD+ data set. They show self shadows and cast shadows across many objects, textures, and materials. The report sums up the methods and results. It gives a benchmark for future shadow removal work.

Catalog record

CREATORFlorin-Alexandru Vasluianu, Tim Seizinger, Zhuyun Zhou, Cailian Chen, Zongwei Wu, Radu TimofteMingjia LiJin HuHainuo WangHengxing LiuJiarui WangQiming HuXiaojie GuoXin LuJiarong YangYuanfei BaoAnya HuZihao FanKunyu WangJie XiaoXi WangXueyang FuZheng-Jun ZhaYu-Fan LinChia-Ming LeeChih-Chung HsuXingbo
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE37 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2506-15524

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

NTIRE 2025 Image Shadow Removal Challenge Report Florin-Alexandru Vasluianu, Tim Seizinger, Zhuyun Zhou, Cailian Chen, Zongwei Wu, Radu TimofteMingjia LiJin HuHainuo WangHengxing LiuJiarui WangQiming HuXiaojie GuoXin LuJiarong YangYuanfei BaoAnya HuZihao FanKunyu WangJie XiaoXi WangXueyang FuZheng-Jun ZhaYu-Fan LinChia-Ming LeeChih-Chung HsuXingbo WangDong LiYuxu ChenBin ChenYuanbo ZhouYuanbin ChenHongwei WangJiannan LinQinquan GaoTong TongZhao ZhangYanyan WeiWei DongHan ZhouSeyed Amirreza MousaviJun ChenHaobo LiangJiajie JingJunyu LiYan YangSeoyeon LeeChaewon KimZiyu FengShidi ChenBowen LuanZewen ChenVijayalaxmi Ashok AralikattiG Gyaneshwar RaoNikhil AkalwadiChaitra DesaiRamesh Ashok TabibUma MudenagudiAnas M. AliBilel BenjdiraWadii BoulilaAlexandru BrateanuCosmin AncutiTanmay ChaturvediManish KumarAnmol SrivastavDaksh TrivediShashwat ThakurKishor UplaZeyu XiaoZhuoyuan LiBoda ZhouShashank ShekharKele XuQisheng XuZijian GaoTianjiao WanSuiyi ZhaoBo WangYan LuoMingshen WangYilin Zhang Abstract This work examines the findings of the NTIRE 2025 Shadow Removal Challenge. A total of 306 participants have registered, with 17 teams successfully submitting their solutions during the final evaluation phase. Following the last two editions, this challenge had two evaluation tracks: one focusing on reconstruction fidelity and the other on visual perception through a user study. Both tracks were evaluated with images from the WSRD+ dataset, simulating interactions between self- and cast-shadows with a large number of diverse objects, textures, and materials. † † footnotetext: \dagger Florin-Alexandru Vasluianu, Tim Seizinger, Zhuyun Zhou, Cailian Chen, Zongwei Wu, and Radu Timofte are the NTIRE 2025 challenge organizers. The other authors participated in the challenge. Appendix A (in the supplementary) provides names and affiliations. https://cvlai.net/ntire/2025 1 Introduction Shadow removal can be considered as a particular sub-task towards the broader field of Image Restoration, having gained significant research interest due to the complexity of the shadow formation process. The intensity and shape of shadows are influenced by a large number of factors, such as the light source, the geometry of the occluding object, and the properties of the recipient surface. The intricacy of this system demands comprehensive investigation, with Shadow Removal and detection persisting as an area of active research. Early works rely on the physical properties of a determined shadow formation model, which is usually characterized by a restricted set of conditions. One category of Shadow Removal solutions using classical image processing aims to restore shadow-free images by transferring local statistics from non-shadowed to shadowed areas finlayson2002removing ; Finlayson_entropyminimization . While promising, such methods require high-quality shadow detection as a pre-processing task to define shadow areas, which constitutes an equally difficult problem. However, solutions built on localization information regarding the shadow affected areas remain a popular design choice 10.1145/1243980.1243982 ; Shor:2008:TSM ; 7893803 . Powered by increasingly faster hardware, later solutions employed deep learning paradigms to exploit large-scale image databases for shadow removal and/or detection Gryka2015softShadows ; SRDDESHADOW ; ISTDwang2018STCGAN ; USRhu2019mask ; vasluianu2023shadow . These solutions are adopting fully-supervised ISTDwang2018STCGAN ; Le_2019_ICCV ; Le_2020_ECCV ; SRDDESHADOW ; zhang2022spa ; vasluianu2023shadow ; guo2023shadowformer , weakly-supervised or self-supervised USRhu2019mask ; ding2019argan ; vasluianu2021shadow ; jin2021dc training frameworks. Recently, diffusion models ho2020denoising have been increasingly adopted by various image shadow removal solutions guo2023shadowdiffusion ; jin2024des3 . These approaches are characterized by the highest perceptual quality in the currently established benchmarks, such as ISTD ISTDwang2018STCGAN , ISTD+ ISTDwang2018STCGAN ; Le_2020_ECCV , or SRD SRDDESHADOW . All aforementioned solutions advanced shadow removal, but the shadow formation model studied by them is limited to specific and simplified scenarios. LRSS Gryka2015softShadows offers soft shadow images, while SRD SRDDESHADOW and ISTD ISTDwang2018STCGAN are limited to scenarios with hard shadows. Their acquisition setups are defined with natural sun light as the only light source, with an occluding object that itself is not visible in the scene, occluding parts thereof in the input image. The reference image is then obtained by simply moving the light occluder away from the scene so that no shadow is cast onto it. Due to the duration of this procedure and variability of natural light due to, e.g. moving clouds, these pairs can show significant variance in color balance. Although a color correction method was introduced in Le_2020_ECCV , defining the improved variations of ISTD+ and SRD+, various semantic inconsistencies still exist in these datasets vasluianu2021shadow ; jin2021dc . The main limitation of the setup used to acquire the aforementioned datasets Gryka2015softShadows ; SRDDESHADOW ; ISTDwang2018STCGAN is that the areas affected by the shadows are always flat with little geometric complexity. Naturally, this is because more complex scene geometries would result in self-shadowing being included in the reference frame. However, self-shadows represent the most prevalent form of natural shadows; thus, their study is imperative for practical applications beyond scenario-specific prototypes. To alleviate these limitations, the authors of WSRD vasluianu2023shadow adopt a professional photography studio setup, using artificial light sources for both input and reference acquisition. While a directional point light source is used to simulate natural sunlight that affects the scene, a omnidirectional softbox-based light setup was used to acquire reference images. This setup enables near-optimal lighting distribution, even when rough surfaces or complex scene geometry are present. Therefore, WSRD can represent complex self-shadowing phenomena, narrowing the gap between the dataset and the in-the-wild domains. Recently, the WSRD setup was extended vasluianu2024towards , with multiple directional lights of varying sizes, positions, and distance. This more complex setup thereby is modeling varying intensities in the umbra-shadow region and varying transitions within the penumbra between shadow and non-shadow affected areas. This advancement broadens the scope of the shadow-removal task by introducing an image database specifically to address the connected Ambient Lighting Normalization problem. This challenge is one of the NTIRE 2025 1 1 1 https://www.cvlai.net/ntire/2025/ Workshop associated challenges on: ambient lighting normalization ntire2025ambient , reflection removal in the wild ntire2025reflection , shadow removal, event-based image deblurring ntire2025event , image denoising ntire2025denoising , XGC quality assessment ntire2025xgc , UGC video enhancement ntire2025ugc , night photography rendering ntire2025night , image super-resolution (x4) ntire2025srx4 , real-world face restoration ntire2025face , efficient super-resolution ntire2025esr , HR depth estimation ntire2025hrdepth , efficient burst HDR and restoration ntire2025ebhdr , cross-domain few-shot object detection ntire2025cross , short-form UGC video quality assessment and enhancement ntire2025shortugc ; ntire2025shortugc_data , text to image generation model quality assessment ntire2025text , day and night raindrop removal for dual-focused images ntire2025day , video quality assessment for video conferencing ntire2025vqe , low light image enhancement ntire2025lowlight , light field super-resolution ntire2025lightfield , restore any image model (RAIM) in the wild ntire2025raim , raw restoration and super-resolution ntire2025raw , and raw reconstruction from RGB on smartphones ntire2025rawrgb . 2 Challenge Data The WSRD dataset vasluianu2023shadow , used in the NTIRE 2023 Image Shadow Removal Challenge vasluianu2023ntire , marked a significant advancement in studying complex shadow interactions by incorporating more intricate surfaces and diverse content. However, its main drawback stemmed from the fact that the data capturing setup is partly mobile. Therefore, inconsistencies in the alignment of pixels have proved to consume significant efforts on the participants side, with complex strategies that alleviate these issues dominating the competition. To address this, we implemented a double-stage image alignment strategy, in which the preliminary image alignment implements a keypoint matching system for homography estimation. Differences in illumination for severely underexposed image segments prove to contribute to a challenging environment for a small subset of outlier image pairs. Therefore, we introduced a refined image alignment as a gradient descent affinity estimation based on object contour geometry as a second alignment step, further enhancing the quality of

Read the complete text