Introduction
The 17th European Conference on Computer Vision (ECCV 2022), held in Tel Aviv, Israel, served as a premier venue for general advancements in the field. However, the accompanying workshops provided a critical platform for the "long tail" of computer vision—specialized domains, emerging challenges, and interdisciplinary applications. This volume (Part I of the proceedings) archives the peer-reviewed contributions from the first four workshop tracks, highlighting innovations in space exploration, artistic analysis, adversarial defense, and autonomous navigation.
Organizational Methodology and Challenges
The coordination of ECCV 2022 workshops was a massive international undertaking, particularly given the global uncertainty of the COVID-19 pandemic. The organizing committee evaluated 103 workshop proposals, ultimately selecting 60 to ensure a balance between established research series and novel, high-potential topics.
To accommodate varying travel restrictions, a hybrid organizational model was adopted:
- Hybrid Format: 30 workshops allowed for both in-person and remote participation.
- Online-Only: 30 workshops were conducted purely virtually.
This volume specifically captures the output of 38 workshops that opted for formal publication, organized into thematic tracks to facilitate cross-pollination of ideas.
Thematic Tracks and Research Highlights
Part I of the proceedings focuses on four primary domains:
1. AI for Space (W01)
Research in this track addresses the unique challenges of extraterrestrial environments. Key focus areas include spacecraft pose estimation, autonomous landing, and mapping. Notable contributions include the use of unsupervised domain adaptation for spacecraft and the application of Neural Radiance Fields (MaRF) for high-fidelity Mars terrain mapping.
2. Vision for Art (W02)
This track explores the "subjective turn" in computer vision, applying algorithms to non-natural image domains. Researchers investigated style transfer, cultural heritage preservation, and the transferability of Vision Transformers (ViTs) to the stylistic nuances of paintings and prints.
3. Adversarial Robustness (W03)
Moving beyond theoretical mathematical perturbations, this track emphasizes "real-world" robustness. Papers explore black-box scenarios and adversarial attacks on complex tasks like object detection and multi-person tracking in dynamic environments.
4. Autonomous Vehicle Vision (W04)
Focusing on the perception stack of self-driving systems, this track covers LiDAR segmentation, pedestrian intention prediction, and gesture recognition. Innovations like "BlindSpotNet" aim to enhance safety by inferring information in areas where sensors have limited visibility.
Key Datasets and Benchmarks
The workshops introduced several new datasets designed to push the boundaries of specialized vision tasks:
- CubeSat-CDT: A cross-domain dataset for 6-DoF trajectory estimation of spacecraft.
- DEArt: A specialized dataset for European Art analysis.
- ARPaS: A benchmark for Adversarially Robust Panoptic Segmentation.
- JackRabbot: A dataset focused on human body pose and navigation within human-centric environments.
Insights and Future Directions
A recurring theme across these workshops is the transition of computer vision into "non-cooperative" and extreme environments. Whether it is estimating the pose of a non-responsive satellite or interpreting the stylistic intent of a 17th-century painting, the field is moving toward more nuanced and robust models. Furthermore, the early integration of Large Language Models (LLMs) for Visual Question Answering in cultural contexts suggests a growing trend toward multimodal foundation models in niche fields.
The editors emphasize that future progress will depend heavily on strengthening academia-industry collaboration to facilitate the technology transfer and scalability of these specialized innovations.
Summary updated