home   program  

Interactive Labeling and Data Augmentation for Vision

Program

Schedule (All times are Eastern)

• 13:00-13:15    Welcome and opening remarks
• 13:15-13:45    Keynote 1: Gim Hee Lee, Deep Spatial Understanding with Fewer Annotated Data - video
• 13:50-14:20    Keynote 2: Angela Dai, Learning to Generate 3D Scenes from Imperfect Data - video
• 14:25-15:30    Poster session - playlist
• 15:30-16:00    Keynote 3: Vittorio Ferrari, Recent works on human-machine collaborative annotation and transfer learning - video
• 16:00-17:00    Poster session (continued) - playlist
• 17:00-17:30    Paper awards and closing remarks

Keynotes

• Vittorio Ferrari
  Recent works on human-machine collaborative annotation and transfer learning
  Summary: I will start by presenting Localized Narratives, a new form of multimodal image annotations connecting vision and language, where annotators describe an image using both their voice and the mouse pointer simultaneously. I will then continue with Pointillism, a protocol to provide point annotations useful for seeding segmentation algorithm, which is immediately clear even to untrained annotators. Though a Pointillism annotation campaign on top of Localized Narratives, we produced annotations for about 2600 classes, with an open-world dictionary completed emerging bottom-up from spontaneous annotator behaviour. Finally, I will give an overview of our recent large-scale systematical study of transfer learning for structured prediction tasks. We carried out over 1200 transfer experiments, and derived insights about the impact of image domain, task type, and dataset size on transfer learning performance.


• Gim Hee Lee
  Deep Spatial Understanding with Fewer Annotated Data
  Summary: Despite the success of deep learning in many computer vision-related tasks, it remains extremely challenging for widespread applications due to the requirement of large amounts of annotated data for training. This problem is further aggravated in spatial understanding tasks, where the ground truth annotations are often laborious to obtain. In this talk, I will share our recent works on bridging non co-occurrence with unlabeled in-the-wild data for incremental object detection; few-shot learning for 3D point cloud semantic segmentation; and unsupervised domain adaptation to transfer knowledge from synthetic to real data for animal pose estimation.


• Angela Dai
  Learning to Generate 3D Scenes from Imperfect Data
  Summary: Reconstructing and understanding real-world 3D environments has seen significant progress in recent years, with the introduction of several real-world 3D scan datasets, but remains challenging due to the imperfect nature of the data, which is typically noisy and incomplete. In this talk, I will present a self-supervised formulation for generative learning for 3D reconstruction from incomplete, real-world scans, followed by self-supervision for colored 3D scene generation by differentiable rendering. Finally, to understand the object composition of a 3D scene, we leverage inexact associations with synthetic data to compose priors for jointly learned object completion and part segmentation in a 3D scene.

Accepted Papers (proceedings)

Nuisance-Label Supervision: Robustness Improvement by Free Labels poster
Xinyue Wei (University of California San Diego)*; Weichao Qiu (Johns Hopkins University); Yi Zhang (Johns Hopkins University); Zihao Xiao (Johns Hopkins University); Alan Yuille (Johns Hopkins University)

EdgeFlow: Achieving Practical Interactive Segmentation with Edge-Guided Flow poster
Yuying Hao (Baidu)*; Yi Liu (Baidu Inc.); Zewu Wu (Baidu); Lin Han (New York University); Yizhou Chen (Chongqing Jiaotong University); Guowei Chen (Baidu, Inc); Lutao Chu (Baidu); Shiyu Tang (Baidu.com, Inc); Zhiliang Yu (Baidu); Zeyu Chen (Baidu, Inc.); Baohua Lai (Baidu)

Data Augmentation for Scene Text Recognition poster
Rowel Atienza (University of the Philippines)*

Multi-Domain Conditional Image Translation: Translating Driving Datasets from Clear-Weather to Adverse Conditions poster
Vishal Vinod (Indian Institute of Science)*; Ram K Prabhakar (Indian Institute of Science); Venkatesh Babu RADHAKRISHNAN (Indian Institute of Science); Anirban Chakraborty (Indian Institute of Science)

Using Synthetic Data Generation to Probe Multi-View Stereo Networks poster
Pranav Acharya (UC Santa Barbara)*; Daniel Lohn (University of California, Santa Barbara); Vivian Ross (University of California Santa Barbara); Maya C Ha (UCSB); Alex N Rich (University of California, Santa Barbara); Ehsan Sayyad (UCSB); Tobias Hollerer (UCSB)

Meta Self-Learning for Multi-Source Domain Adaptation: A Benchmark poster
Shuhao Qiu (Beijing University of Posts and Telecommunications); Chuang Zhu (Beijing University of Posts and Telecommunications )*; WENLI ZHOU (BEIJING UNIVERSITY OF POSATS AND TELECOMMUNICATIONS)

Localizing Human Keypoints beyond the Bounding Box poster
Soonchan Park (Eletronics and Telecommunications Research Institute); Jinah Park (KAIST)*

Learning to Localise and Count with Incomplete Dot-annotations poster
Feng Chen (Computer Vision Laboratory - University of Nottingham)*; Michael P Pound (University of Nottingham); Andrew French (University of Nottingham)

Class-Agnostic Segmentation Loss and Its Application to Salient Object Detection and Segmentation poster
Angira Sharma (University of Oxford)*; Naeemullah Khan (University of Oxford); Muhammad Mubashar (LUMS); Ganesh Sundaramoorthi (Kaust); Philip Torr (University of Oxford)

Reducing Label Effort: Self-Supervised meets Active Learning poster
Javad Zolfaghari Bengar (Computer Vision Center)*; Joost van de Weijer (Computer Vision Center); Bartlomiej Twardowski (Computer Vision Center, UAB); BOGDAN RADUCANU (Computer Version Center)

Self-improving classification performance through GAN distillation poster
Matteo Pennisi (University of Catania); Simone Palazzo (University of Catania); Concetto Spampinato (University of Catania)*

Interactive Labeling for Human Pose Estimation in Surveillance Videos poster
Mickael Cormier (Fraunhofer IOSB, Karlsruhe, Germany)*; Fabian Röpke (Fraunhofer IOSB, Karlsruhe, Germany); Thomas Golda (Fraunhofer IOSB, Karlsruhe, Germany ); Jürgen Beyerer (Fraunhofer IOSB)

Object-Based Augmentation for Building Semantic Segmentation: Ventura and Santa Rosa Case Study poster
Svetlana Illarionova (Skolkovo institute of Science and Technology)*; Sergey Nesteruk (Skoltech); Dmitrii Shadrin (Skoltech); Vladimir Ignatiev (Skoltech); Maria Pukalchik (Skolkovo Institute of Science and Technology); Ivan Oseledets (Skolkovo Institute of Science and Technology)

Bounding Box Dataset Augmentation for Long-range Object Distance Estimation poster
Marten Franke (University of Bremen)*; Vaishnavi Gopinath (University of Bremen); Chaitra Reddy (University of Bremen); Danijela Ristic-Durrant (University of Bremen); Kai Michels (University of Bremen)

Weakly-Supervised Semantic Segmentation by Learning Label Uncertainty poster
Robby Neven (KU Leuven)*; Davy Neven (KU Leuven); Bert De Brabandere (KU Leuven); Marc Proesmans (KU Leuven); Toon Goedemé (KU Leuven - EAVISE)

All you need are a few pixels: semantic segmentation with PixelPick poster
Gyungin Shin (University of Oxford)*; Weidi Xie (University of Oxford); Samuel Albanie (University of Cambridge)

InAugment: Improving Classifiers via Internal Augmentation poster
Moab Arar (Tel Aviv University )*; Ariel Shamir (The Interdisciplinary Center); Amit H Bermano (Tel-Aviv University)

Accepted Papers (extended abstracts)

CutDepth:Edge-aware Data Augmentation in Depth Estimation poster
Yasunori Ishii (Panasonic)*; Takayoshi Yamashita (Chubu University)

Mitigating bias with Targeted Data Augmentations poster
Agnieszka Mikołajczyk (Gdańsk University of Technology)*; Michał Grochowski (Gdańsk University of Technology)