iiinova — where your digital self keeps evolving

Tomas Jakab

@tomasjakab.bsky.social

Postdoc at VGG, University of Oxford | Controllable generation and reconstruction of dynamic 3D objects and scenes | www.robots.ox.ac.uk/~tomj

Posts Replies Media

Tomas Jakab @tomasjakab.bsky.social · 12/06/2025

@oxford-vgg.bsky.social

000

Tomas Jakab @tomasjakab.bsky.social · 12/06/2025

5/ Predicting Dual Point Maps enables tasks beyond 3D reconstruction, such as 3D keypoint detection, viewpoint estimation, fitting a 3D skeleton, and animating the reconstructions.

100

Tomas Jakab @tomasjakab.bsky.social · 12/06/2025

4/ It generalizes so well that it handles a wide variety of horse shapes, even those vastly different from the one it was trained on. It goes even further, generalizing to other quadrupeds like cows and sheep—categories it has never encountered during training in this example!

100

Tomas Jakab @tomasjakab.bsky.social · 12/06/2025

3/ This architecture generalizes remarkably well to real-world images of horses and other quadrupeds, despite being trained solely on synthetic data generated from a single 3D horse model.

100

Tomas Jakab @tomasjakab.bsky.social · 12/06/2025

2/ Dual Point Maps simplify these tasks by reducing them to a pixel labeling problem. Starting from strong image features, the method labels each pixel with its canonical coordinates and, based on those, it then labels the pixels with posed coordinates in camera space.

100

Tomas Jakab @tomasjakab.bsky.social · 12/06/2025

1/ 3D geometric pipelines for deformable objects, such as horses, typically involve estimating the camera, reconstructing the 3D shape in a rest pose, and modeling its deformation. These tasks often require complex models 😥

100

Tomas Jakab @tomasjakab.bsky.social · 12/06/2025

We are presenting Dual Point Maps as a #CVPR highlight tomorrow! Learn about our novel, data-efficient representation for 3D/4D deformable objects—an alternative to classical template shape models. 📍🕑 ExHall D, Poster #100, afternoon session 🌍 dualpm.github.io Brief explanation 👇

131