BYR Achieve · 镜像论坛

We are currently hosting a challenge called “Panoptic Scene Graph Generation (PSG) Challenge”, which asks the participants to solve the PSG task: given a complex scene image as the input, the model should interpret the image with several “subject-verb-object” triplets, which should comprehensively cover the relations in the image. The subject / object should be grounded by a pixel-accurate segmentation mask at the same time. The task is based on our ECCV’22 work: Panoptic Scene Graph Generation. The PSG Challenge is jointly sponsored by International Algorithm Case Competition (hosted by Pazhou Lab) and ECCV’22 SenseHuman Workshop (hosted by MMLab@NTU), with an astonishing $150,000 Prize Pool for the winners (fully sponsored by Pazhou Lab). o Champion: 1 Team - 600,000 RMB (~US$90,000) o Second Prize: 2 Teams - 100,000 RMB each (~US$15,000 each) o Third Prize: 5 Teams - 40,000 RMB each (~US$6,000 each) The preliminary round of the challenge will be ended on Oct. 6th. The top 15 teams will be able to join the final round, which will end on around Nov. 25th, (after the CVPR deadline). We notice that the PSG task generally consists of dense prediction (panoptic segmentation) and scene graph generation, which, we found, well match your research interest. That’s why we are reaching out and inviting you to participate in the PSG challenge to win the prize and further advance the development of comprehensive scene understanding. Also, feel free to recommend other colleagues to join in our challenge! Official Info of PSG Challenge: The main page of the PSG challenge is https://www.cvmart.net/race/10349/base. We provide an English version in CodaLab: https://codalab.lisn.upsaclay.fr/competitions/6734. However, the purpose of CodaLab is only to provide the information for non-Chinese participants. To download the dataset or evaluate the model for ranking and prize-winning, the participants should only use the links from the official competition website. To Download PSG Dataset: https://www.cvmart.net/race/10349/dataset For Submission and Ranking: https://www.cvmart.net/race/10349/my-submission How to register: Notice that participants should register from the official competition website via this link to get access to the data / submission protocol. We recommend users to register with “email registration” option, as sometimes phone number may not get verification code. For Chinese users, please fill out the form that is redirected from the verification email. Non-Chinese users should fill out this form after the verification email is received. The form that is redirected from the verification email is not required to be filled. The scanned document (passport) is required to ensure the legitimacy of the final prize awarding. Then the account will be ready in 1-2 days. If you have any questions regarding the registration or the competition itself, please join our slack space and ask us! More links about PSG Task: PSG Paper: https://arxiv.org/abs/2207.11247 PSG Homepage: https://psgdataset.org/ PSG Dataset Interactive Exploration: https://psgdataset.org/explore.html PSG Codebase: https://github.com/Jingkang50/OpenPSG PSG HuggingFace demo: https://huggingface.co/spaces/mmlab-ntu/OpenPSG PSG Replicate demo: https://replicate.com/cjwbw/openpsg PSG Benchmark: https://paperswithcode.com/task/panoptic-scene-graph-generation/ PSG Forum: https://join.slack.com/t/psgdataset/shared_invite/zt-1d14sdkw3-59pJdrp6gLAHuBObPL91qw We hope the task could further promote comprehensive scene understanding, which could empower exciting downstream tasks such as visual reasoning (based on comprehensive perception) and even relation-aware image generation / editing (assisted by scene graphs, and masks).

【AI竞赛】【百万奖金】全场景图生成