To NeRF or not to NeRF: 

A View Synthesis Challenge for Human Heads

Welcome to View Synthesis Challenge for Human Heads @ ICCV2023 

Novel view synthesis from multiview images is an essential task in computer vision which involves generating novel images or video sequences from viewpoints that were not present in the original input data. This technique has gained significant importance in recent years due to its diverse applications in areas such as augmented and virtual reality, 3D object recognition and reconstruction, and autonomous navigation. By synthesizing new views from existing ones, computer vision systems can enhance the visual quality of images, create new perspectives of scenes, and even generate realistic 3D models of objects or environments. Therefore, the development of new and improved methods for novel view synthesis from multi-view images is vital for advancing the field of computer vision and unlocking its full potential. 

The objective of this challenge is to render novel view images from an unseen view using an input pose, when given a set of 22 (seen view) images and their poses, representing 50 subjects in training. Along with the known images, we provide a corresponding set of masks which result from the image undistortion process – indicating the valid pixels in the image. For the participants convenience, we provide the poses in two common NeRF-based data loader formats: blender and LLFF. While the dataset follows the conventional format for NeRF-based frameworks, we welcome not only neural rendering approaches but also conventional 3D vision/graphics or other innovative methods. Participants can use any combination of the provided dataset. For instance, you can develop a generic model that utilizes multiple subject images for the same view, or you can use subject-specific data to train and render subject-specific validation and test view images. We anticipate methods that overcome the current challenges of novel view synthesis without imposing restrictions on the approaches used. 

Please refer to the 'Participation' section for more detail on how to participate the challenge. The challenge is hosted on CodaLab. More content will be added soon,  please stay tuned :D

Toy Example of the Imperial Light-Stage Head (ILSH) Dataset 

Workshop Program

Oct 2, 2023, Paris, France

Venue: Paris Convention Centre,  Pavilion 7, Hall 7.3, @room W06 

13:30-13:45    Opening remarks and challenge overview presentation


Invited Talks from Academia and Industry


13:45-14:45    Matthias Nießner

14:45-15:45    Thabo Beeler

15:45-16:00    Short Break

16:00-17:00    Fernando De la Torre


Top-ranked Solutions from the Challenge


17:00-17:21    Challenge winner presentation

17:21-17:35    Additional Baseline presentation

17:35-17:50    Award and close remark


News


Important Dates: 

Note: We will announce the final ranking before the workshop dates. Ranking appearing in the leaderboard will be a temporary ranking, as this can be changed after examining all the code and results by organizers. There will be no more extensions for the result submission.

Workshop Organizers

Youngkyoon Jang

Imperial College London

Jiali Zheng

Imperial College London

Jiankang Deng

Imperial College London


Ales Leonardis

University of Birmingham

Stefanos Zafeiriou

Imperial College London

Challenge Organizers

Zheng Li, Keshuang Li, Jifei Song, Helisa Dhamo, Yiren Zhou, Eduardo Perez Pellitero, Christos Kampouris, Athanasios Papaioannou, Alexandros Lattas, Baris Gecer, Foivos Paraperas Papantoniou, Stylianos Moschoglou, Stylianos Ploumpis, Efstathios Galanakis, Rolandos Alexandros Potamias, Wei Zhang, Youliang Yan, Songcen Xu, Xiaofei Wu, Zhensong Zhang, Di Xu, Minglei Li, Haozhe Jia, Changpeng Yang, Zonghong Dai