Maties Machine Learning (MML) is a seminar series and discussion forum with the goal of bringing together people working on machine learning at Stellenbosch University. We meet roughly every second week for short talks on people’s current work, some ML-related topic, or open discussions. The idea is to get to know what others are working on and to strengthen machine learning research at Stellenbosch.

Upcoming talks

5 August, 13:00-13:50 via MS Teams (link will be emailed)

Amir Patel (UCT) - Biomechanics Unleashed: 3D Markerless Motion Capture of Cheetahs in Wild

Animals are capable of incredible agility. However, researchers are still far from understanding these complex dynamics, which have ecological, biomechanical, and evolutionary impacts because they require whole-body motion to quantify. Specifically, the cheetah (acinonyx jubatus) is the pinnacle of manoeuvrability, yet we do not have a complete understanding of its locomotion during hunting in the wild.

Obtaining whole-body kinematic data from wild animals remains an open problem as this data is generated in a controlled lab environment. A further constraint is that these rapid manoeuvres are typically unplanned or require a large area to perform (eg. stopping or turning suddenly). Wildlife videographers however are adept at obtaining incredible footage of animals (like the cheetah) during hunting, but this footage cannot be used for kinematic analysis, as 3D reconstruction requires depth information. Additionally, popular GPS and Inertial Measurement Unit (IMU) collars cannot provide information about whole-body articulation (legs, spine, head and tail) as they treat the animal as a single lumped rigid body.

In this talk, I will discuss recent developments by my lab towards studying biomechanics in the wild. Particularly, I will describe our new dataset (AcinoSet), the world’s first 3D pose dataset of a fast- moving animal in the wild. Using this dataset, we have also developed a method of multi-camera 3D reconstruction (Full Trajectory Estimation - FTE) which is much more accurate than the baseline Sparse Bundle Adjustment (SBA) method. I will also present some unpublished work on monocular 3D whole- body pose estimation of cheetahs.

20 August, 13:00-13:50 via MS Teams (link will be emailed)

Student spotlight talks about speech and vision:

  • Herman Kamper - A quick intro
  • Benjamin van Niekerk - Analyzing speaker information in self-supervised models to improve zero-resource speech processing
  • Leanne Nortje - Direct multimodal few-shot learning of speech and images
  • Kayode Olaleye - Attention-based keyword localisation in speech using visual grounding
  • Christiaan Jacobs - Multilingual transfer of acoustic word embeddings improves when training on languages related to the target zero-resource language

8 September, 13:00-13:50 via MS Teams (link will be emailed)

JC Schoeman - Title TBA

17 September, 13:00-13:50 via MS Teams (link will be emailed)

Arnu Pretorius (InstaDeep) - Title TBA

22 October, 13:00-13:50 via MS Teams (link will be emailed)

  • Mkhuseli Ngxande (20 min) - Title TBA
  • James de Villiers (10 min) - Title TBA
  • Micheal Deyzel (10 min) - Title TBA