List of Accepted Tutorials

Tutorial-1: Introduction to 3D Human Digitization[More Details]




Abstract: Digitizing humans with utmost realism is the holy-grail of Metaverse kind of immersive platform enabling a large set of tele-presence applications, namely, digital gaming, sports analytics, content creation for multimedia/ animation, 3D virtual try-on, etc. This is a challenging task as the body shape geometry evolves over time, yielding a large space of complex body poses as well as shape variations. In addition to this, there are several other challenges such as self-occlusions by body parts, obstructions due to free form clothing, background clutter (in-the-wild setup), sparse set of cameras with non-overlapping fields of views (multi-view setup), sensor noise, etc. Traditionally, image-based techniques for 3D human digitization uses stereo/multi-view (including RGB and depth cameras) setup that typically require studio environments with controlled lighting and multiple synchronized and calibrated cameras. With the advent of learning based methods in 3D Computer Vision, in-the-wild human digitization has become possible. The models such as SCAPE and SMPL models the human 3D surface by parameterizing the body shape and the 3D joint locations and orientation. Model based reconstruction techniques fail to capture accurate geometrical information over the body surface (both body parts and garments) is not retained and are typically applicable only for tight clothing scenarios. In this tutorial, we plan to introduce the following fundamentals to the students/researchers working in the field of 3D computer vision, computer graphics, deep learning, etc.
 -Sensing and 3D representation of the human body
 -Parametric 3D Human model fitting using images, monocular videos, RGB-D sequence, etc.
 -Digitization of clothed human body.
 -Current Research challenges and applications of 3D human digitzation.



Organizers:
Rajendra Nagar
(IIT Jodhpur)
Avinash Sharma
(IIT Jodhpur)



---

Tutorial-2: Photometric Stereo: Past, Present, and Future [More Details]

Abstract: Passive depth sensing in Computer Vision is an active research area as it has benefits over active depth sensors such as less cost, less power consumption, portability, and more depth range. These benefits, however, come at the expense of increased computation. While modern geometric passive techniques based on multi-view stereo and structure from motion have achieved faithful 3D shape reconstruction of objects and large-scale scenes, they fail to model the underlying surface reflectance, making the reconstruction “3D incomplete”. Photometric methods have excelled at recovering the thinnest geometric variation with superior quality and inferring the complete 3D information by jointly estimating the shape, material, and reflectance of the surface. Unlike the geometric methods which perform well only for textured opaque surfaces, photometric methods apply even to transparent, translucent, and textureless objects. Photometric Stereo spans applications such as industrial inspection, medical imaging, cultural heritage preservation, material characterisation, robotics, and entertainment.

This tutorial will introduce the Photometric Stereo method to recover the underlying 3D surface from a series of images captured with a fixed camera, but with varying illumination. This tutorial will cover a thread of deep learning-based photometric methods emphasising the need for uncalibrated, un/self-supervised settings. It will highlight some key applications through recent advances and discuss interesting future research directions for the community to work on.

Organizers:

Shanmuganathan Raman
(IIT Gandhinagar)
Ashish Tiwary
(IIT Gandhinagar)



---

Tutorial-3: Data-free Knowledge Extraction from Deep Neural Networks[More Details]

Abstract: Data-free Knowledge Extraction (DFKE) refers to the process of extracting useful information from a trained deep neural network (DNN) without accessing the underlying training data over which the DNN is trained. The extracted information can be diverse. For instance, it can be a replica of the DNN itself, some sensitive information about the underlying training data, or patterns from thereof, etc. DFKE can be extremely vexing particularly in deployments like MLaaS (Machine Learning as a service). Considering the amount of data, human expertise, and computational resources that are typically required to learn the sophisticated DNNs, it is natural to consider them as intellectual property. Therefore, they need to be protected against any such attempts of extraction (referred to as attacks). On the other hand, philosophically it would be interesting to (i) understand the utility of these trained models without their training data, and (ii) formulate guarantees on the information leakage (or extraction). In this tutorial, I plan to first introduce the phenomenon of data-free model extraction and discuss different ways in which it can be manifested, both in white-box and black-box scenarios. In the later part, I will focus more on the potential threats of leaking sensitive information about the training data to a dishonest user in the form of different attacks. Finally, I will discuss some of the active directions to investigate further.


Organizers:

Konda Reddy Mopuri
(IIT Hyderabad)



---

Tutorial-4: MEDical imaging - Data-EffiCIent machine LEarning - (MED-DECILE)[More Details]

Abstract:The extraordinary outcomes of deep learning methods should be viewed in the light of certain prerequisites, the first and foremost being the necessity of large amounts of clean labeled data. For medical imaging, dataset labeling costs are pretty steep given the domain expertise required for labeling medical data. This, in combination with the high computing cost that is associated with deep learning algorithms, prevents the democratization of utilizing machine learning algorithms in low- and middle-income countries. This is especially true for India which is data-rich due to the rapid digitalization of medical imaging procedures but labeled-data poor. In this tutorial, we will demonstrate several robust ML optimization techniques that can be used in low-data, low-resource settings. Our tutorial contains several semi-supervised and weakly supervised methods that utilize shallow machine-learning models and heuristic-driven models as labeling functions to label only a subset of unlabeled data points, instead of the whole set. We will demonstrate how close are these methods to the state-of-the-art deep learning methods, demonstrating in their utility for labeling medical datasets.

Organizers:

Ganesh Ramakrishnan
(IIT Bombay)
Venkatapathy Subramanian
(IIT Bombay)
Kshitij Jadhav
(IIT Bombay)

---

---