ALLEGRO/GAYA workshop


Venue: A103 (before coffee), Grand Amphithéâtre (after coffee), Inria Grenoble - Rhône-Alpes (Montbonnot/Inovallée site: Directions)

Tuesday, 5th June

09:15 - 10:00   Thomas BroxLearning motion and 3D perception [slides]
TBD
10:00 - 10:45   Iasonas KokkinosDensePose: Dense Pose Estimation In The Wild Abstract [slides (pdf)]
In this talk I will present recent work on establishing dense correspondences between 2D images and 3D surface models "in the wild", namely in the presence of background, occlusions, and multiple objects. I will start by describing DenseReg and DensePose, two recently introduced systems for this goal which operate at multiple frames per second on a single GPU. Time permitting I will cover more recent works on extending such techniques to unsupervised learning and combining them with structured prediction.
10:45 - 11:15   coffee
11:15 - 12:00   Jean PonceDeformable kernel networks for joint image filtering and Musings on word embeddings [slides (pptx)]
12:00 - 12:45   Nicolas MansardDevelopment of a memory of motion for locomotion of legged robots on 3D terrains Abstract [slides]
For years, our team developed a complete approach for motion generation and control of a humanoid robot locomoting in 3D environments. Our work is deeply based on optimal control, and more specifically its use in trajectory optimization. The motion is formalized as a numerical optimization problem, that we try to solve on-line at high frequency: a motion problem is typically nonlinear, with ~5000 variables (but some sparsity) and should be solved 10 to 100 times per seconds. To guide the optimizer, we developed several dedicated motion models, of smaller dimension, hence faster to solve. These additional models are somehow ad hoc and heavily dependent on the application context (e.g. they can be used in locomotion, but would not be able to generalize to crawling, dexterous manipulation, etc). In the presentation, I will explain our general approach, describe the motion models that we developed, and describe how we attempt to generalize these dedicated developments used data-based approaches, by formalizing a "memory of motion".