Computer Vision - ECCV 2022 Workshops - Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VIII

Leonid Karlinsky, Ko NishinoECCV 2023 · DOI · DBLP

Introduction

The 17th European Conference on Computer Vision (ECCV 2022) hosted a series of workshops aimed at addressing the "last mile" problems in computer vision. While primary conference tracks often prioritize general improvements in core tasks like object detection and classification, these workshops focus on the transition from laboratory performance to real-world reliability. This volume, Part VIII of the proceedings, captures the community's progress in high-stakes domains including safety-critical uncertainty quantification, specialized human communication, and precise spatial intelligence.

Specialized Domains and Real-World Challenges

The research presented tackles the gap between controlled datasets and the unpredictability of the physical world. Key focus areas include:

  • Reliability in Safety-Critical Tasks: Developing models that can quantify their own uncertainty, providing a signal when they are "unsure" of a prediction.
  • Domain Shift and OOD Generalization: Ensuring robustness when models encounter data distributions different from their training sets.
  • Sign Language Understanding: Moving beyond simple gesture recognition to capture the complex, continuous linguistic structures of sign languages.
  • 6D Spatial Intelligence: Advancing from 2D bounding boxes to a full 6D understanding of an object's position and orientation.

Several recurring themes emerge across the diverse methodologies showcased in this volume:

  • The Rise of Transformers: Attention-based architectures have expanded beyond natural language processing to become the backbone for tasks like 6D pose estimation (e.g., Trans6D) and state-of-the-art visual tracking.
  • Integration of Geometric Priors: Despite the dominance of deep learning, successful methods like CenDerNet re-incorporate traditional geometric concepts, such as curvature and 3D rendering, to enhance spatial accuracy.
  • Self-Supervised Learning: To combat the high cost of labeling in specialized fields like medical imaging and depth estimation, researchers are increasingly leveraging self-supervision on unlabeled data.
  • Multimodal Fusion: Complex tasks like visual grounding—linking natural language to specific image regions—now rely on the integration of text, video, and 3D data.

Key Results and Benchmarks

The workshops are anchored by influential challenges that drive the state-of-the-art:

  • VOT2022 Tracking: The 10th anniversary of the Visual Object Tracking challenge demonstrated that hierarchical cross-attention transformers are the current state-of-the-art, significantly outperforming previous correlation-filter-based methods.
  • 6D Pose Refinement: The Trans6D approach achieved significant gains in object pose refinement, proving that attention mechanisms can effectively capture complex spatial relationships.
  • Sign Language Spotting: New multi-modal fusion techniques showed a marked improvement in identifying specific signs within unconstrained, continuous video streams compared to single-modality baselines.
  • OOD Generalization: Challenges focused on compositional and multimodal perception highlighted the ongoing difficulty of maintaining performance across domain shifts.

Discussion and Key Insights

A major takeaway from this collection is that uncertainty is now a first-class citizen. For computer vision to be viable in robotics or medicine, models must be "self-aware" regarding their confidence levels. Furthermore, research into abstract imagery and drawings has revealed a significant "natural image" bias in standard models, which often overfit to photorealistic textures and fail when these cues are absent. Finally, there is a clear tension between the massive models required to win challenges and the need for "lightweight" architectures (like YORO) suitable for edge device deployment.

Conclusion and Future Work

While significant strides have been made, the "long tail" of rare objects, unusual lighting conditions, and diverse human behaviors remains a substantial hurdle. Future work must address the computational expense of top-performing models to enable real-time deployment and improve the interpretability of complex architectures like Transformers to understand why they fail in critical scenarios.

Summary updated

BibTeX
@inproceedings{karlinsky2023computer,
  author = {Leonid Karlinsky and Ko Nishino},
  title = {{Computer Vision - ECCV 2022 Workshops - Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VIII}},
  booktitle = {European Conference on Computer Vision},
  year = {2023},
  doi = {10.1007/978-3-031-25085-9},
}

Loading related content...

Figure