Computer Vision – Centro de Investigación y Formación en Inteligencia Artificial | Uniandes

Robustness

admin — Mon, 15 Aug 2022 22:52:34 +0000

Computer Vision systems have achieved remarkable performances across a wide variety of tasks, such as recognition, segmentation, detection, and generation. However, these systems have also been shown to be vulnerable against semantically-meaningless perturbations. In particular, recent works have shown that these systems, while accurate, lack robustness. This property is undesirable for intelligent systems on which we wish to rely on in the real world. In the Center, we have worked on robustness on various dimensions. In particular, we have (1) designed biologically-inspired techniques to improve robustness, (2) proposed novel semantically-oriented dimensions for the assessment of the robustness, (3) studied how inexpensive techniques during system deployment can provide robustness benefits, (4) investigated the pervasiveness of the lack of robustness in the medical domain, and (5) shown how techniques for improving robustness can be harnessed to improve the performance of super-resolution systems.

3D Vision

admin — Thu, 03 Jun 2021 03:23:28 +0000

Encouraged by evolving research fields such as Artificial Reality, Autonomous vehicles, and scene understanding, 3D Vision problems have gained interest recently. Many of the tasks studied in 3D Vision are inspired by its 2D counterpart. However, the extending of deep learning into depth and 3D information can unlock a variety of applications. Contrary to 2D Vision problems that study 2D standard-images, 3D data is structured in different formats, including RGB-D images, voxel grids, point clouds, and meshes. This diversity and the need for low-computational-cost processing is a challenge for research in this area.

Segmentation and Grouping

admin — Thu, 03 Jun 2021 03:18:59 +0000

Image Generation

admin — Thu, 03 Jun 2021 02:47:27 +0000

Generation is the task of creating new data from a specific dataset. Image generation is the line of research that focuses on creating images whose distribution is similar to that of the training dataset. We aim at designing novel methods that might be conditioned on specific attributes of the training dataset to create high-quality images not only for natural images but also for medical images.

Presentation video

Video Analysis

admin — Thu, 03 Jun 2021 02:37:26 +0000

Current media is transitioning from an image-based information to video-based data. This arises the need of computational tools to analyse this kind of information. This project concentrates its efforts to form efficient tools to process video related tasks.