Abstract: These keynotes discuss the following: Computer vision; Pattern recognition and machine learning; Image, speech, signal and video processing; Document analysis, biometrics and pattern recognition applications; Biomedical image analysis.
Introduction
The frontier of computer vision and machine learning is currently defined by a shift from rigid, black-box models toward systems that are more statistically optimal, human-centric, and efficient in their use of data. This collection of keynote summaries addresses five critical bottlenecks in modern AI: geometric estimation noise, the accessibility of machine learning (ML) tools for non-experts, the high cost of fine-grained data annotation, identity variance in surveillance, and the modeling of high-dimensional complexity in biomedical imaging.
Rethinking Geometric Estimation
Kenichi Kanatani challenges the standard practice of cost-function minimization—such as least squares or Maximum Likelihood (ML) estimation—which often proves inaccurate or inefficient when estimating geometric structures from noisy data. He introduces "hyperrenormalization," a non-minimization approach. Instead of searching for a global minimum, the method solves a specific matrix equation. This shift provides superior statistical accuracy and computational efficiency, outperforming traditional Sampson error minimization and establishing a new standard for geometric computer vision.
Bridging the "Expert Gap" with Interactive ML
Mark Girolami identifies a significant barrier to the adoption of AI: while ML tools are powerful, they remain largely inaccessible to scientists in other fields, such as ecology, who lack specialized knowledge in data pipelines. He advocates for "Scientist-in-the-Loop" systems. By moving away from the rigid "label-train-evaluate" standard, Girolami proposes interactive interfaces that allow domain experts to inject their specialized knowledge directly into the learning process, treating the human as a vital feature of the system rather than just a source of labels.
Unsupervised Part Discovery in Fine-Grained Recognition
Fei-Fei Li addresses the "annotation bottleneck" associated with fine-grained recognition—the task of distinguishing between very similar categories, such as specific bird species or car models. Traditionally, this requires expensive, manually annotated part labels. Li’s team utilizes Convolutional Neural Networks (CNNs) to learn appearance descriptors combined with a novel unsupervised method to detect discriminative parts (e.g., a bird’s beak or a car’s headlight). By looking at images with similar poses, the system discovers these parts automatically, achieving high accuracy without the need for costly manual segmentation masks.
Robustness in Person Re-Identification
Stan Z. Li tackles the challenge of "Person Re-Identification," where dramatic changes in lighting, pose, and background make matching individuals across different cameras difficult for traditional hand-crafted features. His approach utilizes a symmetrical "Siamese" network connected by a cosine layer to learn color and texture features simultaneously. A key innovation is the use of binomial deviance as a cost function, which ensures the model remains robust against the outliers and extreme variations common in real-world surveillance footage. This method has demonstrated superior performance on the VIPeR dataset and strong generalization across different databases.
Graph-Based Discrete Optimization in Biomedicine
Nikos Paragios explores the modeling of complex visual perception tasks in the medical field, where data is often high-dimensional and noisy. He proposes casting these perception tasks as graph optimization problems. By utilizing graph-based discrete optimization, researchers can develop more efficient inference algorithms that bridge the gap between theoretical computer vision and the practical requirements of biomedical imaging.
Conclusion
These diverse contributions highlight a unified trend in AI research: the move toward "smarter" algorithms that require less manual intervention and offer greater statistical rigor. Whether through the non-minimization techniques of hyperrenormalization, the unsupervised discovery of object parts, or the integration of human expertise into the ML pipeline, these advancements pave the way for more robust and accessible intelligent systems.
Summary updated