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

Leonid Karlinsky, Ko NishinoECCV 2023 · DOI · DBLP

Abstract: The 17th European Conference on Computer Vision (ECCV 2022) workshop proceedings highlight a critical shift from general-purpose theory to specialized, high-impact applications. This volume focuses on three primary pillars: optimizing deep learning for mobile edge devices, automating civil infrastructure monitoring, and refining AI accuracy for high-stakes medical diagnostics. By addressing the "deployment gap," these workshops showcase how computer vision is evolving to meet real-world resource constraints and environmental complexities.

Introduction: The Deployment Gap

A central challenge in modern computer vision is the discrepancy between performance on massive, general datasets (like ImageNet) and performance in specialized real-world environments. This "deployment gap" is characterized by three main factors:

  1. Resource Constraints: The need for models light enough for mobile phones or embedded sensors.
  2. Environmental Complexity: The requirement for high-precision segmentation in "dirty" data environments, such as construction sites with low light and occlusions.
  3. High-Stakes Variability: The "domain shift" in medical imaging caused by varying equipment and techniques, necessitating robust and explainable models.

Methodological Innovations

The proceedings highlight several novel algorithmic trends designed to overcome these hurdles:

  • Hybrid Architectures: EdgeNeXt represents a new class of models that combine the local feature extraction of CNNs with the global reasoning of Transformers, specifically optimized for mobile vision tasks.
  • Efficiency Techniques: Researchers are exploring Hydra Attention for multi-head efficiency, Post-training Quantization (QFT) to reduce model size without retraining, and N:M fine-grained sparsity to accelerate activations.
  • Self-Supervision & Domain Adaptation: To mitigate the lack of labeled data in niche fields, Self-supervised Pretraining is being used for material classification in mining, while FUSION introduces unsupervised test-time stain adaptation for medical imaging generalization.

Specialized Benchmarks and Datasets

To drive progress in these sub-fields, several new benchmarks were introduced:

  • CrackSeg9k: A large-scale collection specifically for crack segmentation in infrastructure.
  • ConSLAM: A real-world construction dataset designed for testing SLAM (Simultaneous Localization and Mapping) and long-term progress monitoring.
  • Medical Imaging Sets: New research utilizes hyperspectral RGB datasets for building facades and various MRI/CT/Histology datasets for COVID-19 and cancer grading.

Key Insights: Beyond Static Models

  • Multi-Dimensional Efficiency: Efficiency is no longer just about speed; it now encompasses "Power Awareness" and the ability to perform "Continual Inference" on live data streams.
  • Lifelong Learning in Construction: Models must be dynamic. The concept of "Lifelong Site Object Detection" allows AI to adapt as a building progresses from a skeleton frame to a finished structure.
  • Explainability in Medicine: For AI adoption in radiology, "Model-Assisted Labeling via Explainability" is essential to help clinicians understand why specific lung nodules or biopsies were flagged.

Results and Future Directions

Architectures like EdgeNeXt have achieved new state-of-the-art (SOTA) balances between accuracy and latency on mobile benchmarks. In infrastructure, automated UAV-based inspections have shown high reliability. Furthermore, 3D-ConvNeXt and BERT-based features have improved COVID-19 detection and severity grading from CT scans.

Despite these gains, limitations remain. Inter-rater variability among experts makes "ground truth" difficult to define in medical AI, and data imbalance continues to plague civil engineering datasets. Future work is directed toward more robust Unsupervised Domain Adaptation to ensure models can generalize to unseen environments.

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 VII}},
  booktitle = {European Conference on Computer Vision},
  year = {2023},
  doi = {10.1007/978-3-031-25082-8},
}

Loading related content...

Figure