Learning from Limited and Imperfect Data (L2ID)

A joint workshop combining Learning from Imperfect data (LID) and Visual Learning with Limited Labels (VL3)

June 20, 2021 (Full Day, Virtual Online)


  • YouTube channel with videos from speakers, orals, posters, and panel sessions is here
  • Introduction slides are available here
  • The workshop will use the CVPR Zoom link; see the June 20th list of workshops and search for Limited or Imperfect Data. ONLY the poster session at noon PDT will be on Gatherly.
  • YouTube video presentations from invited speakers available here! Please ask questions in the comments section and we will ask them during the panels!
  • Full list of accepted papers and oral papers available!


Learning from limited or imperfect data (L^2ID) refers to a variety of studies that attempt to address challenging pattern recognition tasks by learning from limited, weak, or noisy supervision. Supervised learning methods including Deep Convolutional Neural Networks have significantly improved the performance in many problems in the field of computer vision, thanks to the rise of large-scale annotated data sets and the advance in computing hardware. However, these supervised learning approaches are notoriously "data hungry", which makes them sometimes not practical in many real-world industrial applications. This issue of availability of large quantities of labeled data becomes even more severe when considering visual classes that require annotation based on expert knowledge (e.g., medical imaging), classes that rarely occur, or object detection and instance segmentation tasks where the labeling requires more effort. To address this problem, many efforts, e.g., weakly supervised learning, few-shot learning, self/semi-supervised, cross-domain few-shot learning, domain adaptation, etc., have been made to improve robustness to this scenario. The goal of this workshop is to bring together researchers to discuss emerging new technologies related to visual learning with limited or imperfectly labeled data. Topics that are of special interest (though submissions are not limited to these):

  • Few-shot learning for image classification, object detection, etc.
  • Cross-domain few-shot learning
  • Weakly-/semi- supervised learning algorithms
  • Zero-shot learning · Learning in the “long-tail” scenario
  • Self-supervised learning and unsupervised representation learning
  • Learning with noisy data
  • Any-shot learning – transitioning between few-shot, mid-shot, and many-shot training
  • Optimal data and source selection for effective meta-training with a known or unknown set of target categories
  • Data augmentation
  • New datasets and metrics to evaluate the benefit of such methods
  • Real world applications such as object semantic segmentation/detection/localization, scene parsing, video processing (e.g. action recognition, event detection, and object tracking)

Challenge Information

This year we have two groups of challenges: 1) Localization and 2) Classification. The due date for submission is May 14th, 2021 (Deadline extended to May 28th!).


Workshop Paper Submission Information

The contributions can have two formats
  • Extended Abstracts of max 4 pages (excluding references)
  • Papers of the same length of CVPR submissions
We encourage authors who want to present and discuss their ongoing work to choose the Extended Abstract format.
According to the CVPR rules, extended abstracts will not count as archival.
The submissions should be formatted in the CVPR 2021 format and uploaded through the L2ID CMT Site

Please feel free to contact us if you have any suggestions to improve our workshop!    l2idcvpr@gmail.com


(All times in PT Time Zone)

Date Speaker Topic
8:00-8:05 PT / 11:00-11:05 EDT Organizers Introduction and opening
8:05-8:45 PT / 11:05-11:45 EDT Guoliang Kang - Pixel-Level Cycle Association: Domain Adaptive Semantic Seg.
Angela Dai - Learning from Imperfect RGB-D Scan Data
Colin Raffel - Explicit and Implicit Entropy Minimization in Proxy-Label-Based SSL
Sanja Fidler - Image GANs for Reducing Pixel-Wise Supervision
Oral Papers: [A, B, D, H]
Unlabeled data / Self/Semi-Supervised, Domain Adaptation
8:45-9:15 PT / 11:45-12:15 EDT [A, B, D, H, Classification Challenge Participants] Paper Spotlight Talks
9:15-9:55 PT / 12:15-12:55 EDT Chelsea Finn - Few Shot Learning in the Real World
Rogerio Ferris - How Transferable are Contrastive Representations?
Trevor Darrell - Recent Progress on Unsupervised Detection and Adaptation
Oral Papers: [G, J, L]
Few Shot Learning
9:55-10:10 PT / 12:55-13:10 EDT Coffe Break
10:10-10:50 PT / 13:10-13:50 EDT Boqing Gong - When Vision Transformers Outperform ResNets
Vahan Petrosyan - Tools to share datasets and find imperfect data in CV
Olga Russakovsky - Mitigating bias and privacy concerns in visual data
Dina Katabi - Making Contrastive Learning Robust to Shortcuts and Generalize it to New Modalities
Oral Papers: [E, K]
Robustness, adversarial, bias/fairness, deployment/industry
10:50-11:20 PT / 13:50-14:20 EDT Oral Papers: [G, J, L, E, K] Paper Spotlight Talks
11:20-12:00 PT / 14:20-15:00 EDT Alexander Schwing - Not All Unlabeled Data are Equal
Humphrey Shi - Escaping the Big Data Paradigm with Compact Transformers
Anurag Arnab - Video Understanding with Imperfect Data
Oral Papers: [C, F, I]
Imperfect/Noisy/Weakly supervised
12:00-14:00 PT / 15:00-17:00 EDT Gatherly Poster / Lunch Break
14:00-14:40 PT / 17:00-17:40 EDT Aarti Singh - Learning from preferences and labels
Philip Isola - When and Why Does Contrastive Learing Work?
14:40-15:10 PT / 17:40-18:10 EDT Oral papers: [C, F, I, Localization Challenge Participants] Paper Spotlight Talks
15:10-15:50 PT / 18:10-18:50 EDT All available Future Directions
15:50-16:00 PT / 18:50-19:00 EDT Organizers Wrap-up Discussion
ID Title
A Training Deep Generative Models in Highly Incomplete Data Scenarios with Prior Regularization
B Unsupervised Discriminative Embedding for Sub-Action Learning in Complex Activities
C Unlocking the Full Potential of Small Data with Diverse Supervision
D Distill on the Go: Online knowledge distillation in self supervised learning
E Learning Unbiased Representations via Mutual Information Backpropagation
F PLM: Partial Label Masking for Imbalanced Multi-label Classification
G ReMP: Rectified Metric Propagation for Few-Shot Learning
H A Closer Look at Self-training for Zero-Label Semantic Segmentation
I An Exploration into why Output Regularization Mitigates Label Noise
J Shot in the Dark: Few-Shot Learning with No Base-Class Labels
K Contrastive Learning Improves Model Robustness Under Label Noise
L A Simple Framework for Cross-Domain Few-Shot Recognition with Unlabeled Data


Important Dates

Description Date
Paper submission deadline March 25th, 2021
Notification to authors April 8th, 2021 (extended to Apr 13)
Camera-ready deadline April 20th, 2021
Challenge submission deadline May 14th, 2021