Pose-guided Human Rendering Benchmark Challenge
Current
First phaseEnd
Competition EndsHUMBI Pose-Guided Human Rendering Benchmark Challenge
HUMBI is an ideal dataset to evaluate the ability of modeling human appearance. To measure such ability, we formulate a novel benchmark challenge on a pose-guided appearance rendering task: given a single view image of a person, render the person appearance from other viewsand poses. HUMBI offers the ground truth of this challenging task where the performance of the approaches can be precisely characterized. We validate the feasibility of thebenchmark challenge using the state-of-the-art rendering methods.
Task Definition
- Input: a source person image and target body pose (2D keypoints).
- Output: a synthesized person image that agrees with the taget pose and the source person identity.
Dataset Download
- Training data: download (100 identities, 366,300 images)
image, 2D keypoints, training list text file
- Training data (Mask): download
- Validation data: download (20 identities, 7,674 images, 8,365 validation pairs)
images, foreground mask, 2D keypoints, validation list text file)
- Validation data (Mask): download
- Testing data: download (20 identities, 378 images, 1,079 testing pairs)
image (only for source), 2D keypoints, testing list text file
- Testing data (Mask): download
Dataset Description
- Image: 256x256 cropped person image captured from multiple viewpoints
- Foreground mask: binary mask detected from image using a person segmentation method
- keypoints: a txt file for 2D human body joints made by Openpose
- txt file: the file that includes the paired name for source and target
- Mask: Foregroud mask detected from image using recent instance-wise person segmentation
How to submit?
- Create the all testing results listed in the testing file where the name order should be like "145_051_088_145_147_006.jpg" as described in the testing pair list file.
- Create .zip file inside of the predicted results folder.
- Submit the .zip file in the participate->submit section.
Evaluation Code Example
References and credits:
1) HUMBI: A Large Multiview Dataset of Human Body Expressions and Benchmark Challenge (TPAMI under review) [paper]
Jae Shin Yoon, Zhixuan Yu, Jaesik Park, and Hyun Soo Park
2) HUMBI: HUMBI: A Large Multiview Dataset of Human Body Expressions (CVPR 2020) [paper]
{Zhixuan Yu*, Jae Shin Yoon*}, In Kyu Lee, Prashanth Venkatesh, Jaesik Park, Jihun Yu, and Hyun Soo Park
Project page: HUMBI.
Contact: jsyoon [at] umn.edu
Submission Rules
1. You may submit 3 entries every day.
2. You cannot sign up multiple accounts and therefore you cannot submit from multiple accounts.
Training Rules
1. You are "not" allowed to use any testing data in training time.
2. You are "not" allowed to use any external data.
3. You are allowed to use validation set in training time and freely.
4. You are allowed to edit the training and validation .txt files (e.g., add more training pairs).
5. You are allowed to augment dataset (e.g., translation, rotation, and so on) within the provided training and validation sets.
6. You are allowed to use additional detection results (e.g., fashion segmentation densepose detection, single-view depth prediction, and so on) within the dataset.
Evaluation Rules
1. Each submission will be scored and ranked by the evaluation metrics stated on the Competition Website. The ranking will be visible on the Competition Website's leaderboard if the submission is set to the public. The potential winner(s) are determined solely by the leaderboard ranking on the leaderboard, subject to compliance with these Rules.
2. In the event of a tie, the Submission that was entered first to the Competition will be the winner.
3. The ranking will be based on the MSSIM
First phase
Start: June 30, 2013, midnight
Competition Ends
Never
You must be logged in to participate in competitions.
Sign In| # | Username | Score |
|---|---|---|
| 1 | cpmolnar | 0.829868 |
| 2 | humbi | 0.812742 |
| 3 | ejmccalla | 0.813914 |