Join us at #ICCV2023 in Paris for the inaugural workshop on Robust Computer Vision Across Geographies, which includes keynotes by an exciting line-up of speakers, and oral talks by the winners of the GeoNet domain adaptation challenge! Website: geonet-challenge.github.io/I…

🗓️ 10/27 Poster 27 at 11:00 AM (Hall B). MemSAC: Memory Augmented Sample Consistency for Large Scale Domain Adaptation @tarun_05 @AstutiSharma @manu_chandraker Cross-domain sample consistency approach to scale domain adaptation to large-scale datasets. tarun005.github.io/MemSAC/

🗓️ 10/25 Oral at 1:30 PM (Hall E). Poster 115 at 3:30 (Hall B). Physically-Based Editing of Indoor Scene Lighting from a Single Image. Zhengqin Li, @manu_chandraker et. al. Estimate visible and invisible light sources for indoor scene relighting ! bit.ly/3MY8TaB

🗓️ 10/25 Oral at 1:30 PM (Hall E). Poster #80 at 3:30 PM (Hall B). A Level Set Theory for Neural Implicit Evolution under Explicit Flows @ishit_mehta @manu_chandraker Ravi R. A level-set theory bridge between explicit and implicit shape representations. ishit.github.io/nie

The key insight is that gradients for energy minimization on explicit surfaces behave as explicit flow fields acting on the 0-surface of level-sets. This allows extensions to classical level-set theory to optimize for mesh-based energy functions with neural implicits.

Our theory of neural implicit evolution bridges an existing gap between Lagrangian shape representations like triangle meshes and Eulerian representations like parametric level-sets. It explains and improves recent works on differentiable rendering and iso-surface extraction.

We demonstrate applications such as inverse rendering with high-genus recovery, curvature-based deformations for surface smoothing and mean-curvature flow, and ARAP editing on neural implicits.

Check out our new work on level-set evolution of neural implicit surfaces. It allows energy formulations on triangle meshes to be used to deform neural implicits. A variety of explicit differentiable rendering and surface algorithms can now be used with free topological changes.

Excited to share #PhotoScene at #CVPR2022 for creating photorealistic 3D digital twins from images! Chat with us (in-person!!) at poster session 4.1 during 10:00-12:30 on June 24 (Fri) and see our project page for more details! Project page: yuyingyeh.github.io/projects…

Excited to invite you all to the oral talk (session 1.2, 1:30 PM, in-person!) of our paper #IRISFormer at #CVPR2022, where we propose a dense #transformer for single image inverse rendering. Also find us at the poster session 1.2 to learn more about #IRISFormer and chat with us!

#OpenRooms also provides ground truth semantics and physical properties such as friction coefficients. Integration with physics engines such as PyBullet allows robotics applications like rearrangement.

#OpenRooms is the first dataset that provides high-quality ground truth for photorealistic AR and VR, including SVBRDF and SV-lighting with direct, indirect and visibility components.

Check out #OpenRooms being released by @ViLabUCSD at #CVPR2021, an open framework for creating photorealistic counterparts of indoor 3D scans, with ground truth geometry, material, lighting, semantics and physical properties.

Using this idea, we improve over prior adversarial approaches on multiple datasets for domain adaptation such as Office-31 and Birds-31, and show that ILA-DA can be used complementary to any adversarial adaptation algorithm.

ILA-DA uses a nearest neighbor based target labeling, followed by a multi-sample contrastive loss to re-align positive and negative samples from different categories in the target dataset.

In contrast, the novelty of ILA-DA is that it uses pairwise similarity affinities across labeled source and unlabeled target examples which is more robust to outliers and noisy classifier boundaries!

Many global adversarial domain alignment methods suffer from negative transfer, while class aware adaptation methods are sensitive to noisy pseudo-labeling, affecting target accuracy.

Unsupervised Domain Adaptation (UDA) involves training deep learning models that are robust to test-time distribution shifts, using only unlabeled target data during training.