πŸš€
1st CLEAR Challenge (CVPR'22)
CVPR 2022 Open World Vision Workshop Challenge Track
The 1st CLEAR challenge was hosted on CVPR'22 Open World Vision workshop in June 19th, 2022. Through the challenge, we hope to encourage the continual learning (CL) community to embrace real-world CL problem by:
  • Working on scenarios with natural distribution shifts, e.g., CLEAR-10 and CLEAR-100.
  • Addressing the generalization bottleneck of practical CL systems, especially the train-test domain gap and forward transfer.
  • Not limiting the replay buffer size (unlike in most prior works) because memory cost is usually not the bottleneck. Instead, we use a more realistic proxy of total training cost, that is to limit the training time per bucket to be 12hr (on a single 2080ti, see this chart for conversion to other types of GPU).
At the end of the challenge, we have 79 participants from 21 different countries and region joining the challenge, and together they pushed the state-of-the-art on CLEAR-10/CLEAR-100 to another level.
Demographics for 1st CLEAR challenge
Top performing teams in this challenge adopt various stratgies to improve the generalization bottleneck, such as:
  • Experience Replay for more efficient usage of diverse training samples.
  • Data augmentation, e.g., Cutmix, Mixup, etc., for improving data efficiency.
  • Special losses to improve generalization, e.g., sharpness aware minimization, contrastive-based representation learning, etc.
Please checkout the slides for a quick summary of the workshop.

Workshop result

Top-4 and Innovation Prize on 1st CLEAR challenge
1st Place (Youtu Lab) summary video (link to repo):
shennong3_clear.mp4
105MB
Binary
1st Place Summary Video
2nd Place (AI Prime) | solution code​
Copy link