Head over to our poster or visit our website to learn more about Deblur e-NeRF, as well as the event camera simulator & dataset released alongside. 🧵[4/4]

In this work, we propose a physically-accurate pixel bandwidth model that accounts for event motion blur, which we incorporate for blur-minimal NeRF reconstruction from motion-blurred events, generated under high-speed or low-light conditions. 🧵[3/4]

Contrary to what most think, event cameras also suffer from motion blur, especially under high-speed or low-light conditions. However, no prior work on reconstructing NeRFs from events, nor event simulators, have considered the full effects of event motion blur. 🧵[2/4]

Thrilled to share my #ECCV2024 work with Gim Hee Lee (@gimhee_lee) on "Deblur e-NeRF: NeRF from Motion-Blurred Events under High-speed or Low-light Conditions" tomorrow morning at Poster Session 1 (Poster #248)! Project page: https://t.co/TFq2KwvqrG 🧵[1/4]

I'll be presenting Robust e-NeRF, our work on robust NeRF reconstruction with moving event cameras, at #ICCV2023 this Friday morning (Poster 071, Foyer Sud). Drop by to learn more and have a chat! Check out our thread below too!

Check out the website to learn more about Robust e-NeRF, as well as the event camera simulator & dataset released alongside. Do also drop by the "3D from Multi-View and Sensors 2" @ FRI-AM poster session (10.30am - 12.30pm, 6 Oct 2023) to have a chat! 🧵[7/N=7]

2. A complementary pair of normalized reconstruction losses that can effectively generalize to arbitrary speed profiles and intrinsic parameter values without such prior knowledge 🧵[6/N]

Robust e-NeRF achieves this with 2 key components: 1. A realistic event generation model that accounts for various intrinsic parameters (e.g. time-independent, asymmetric threshold & refractory period) and non-idealities (e.g. pixel-to-pixel threshold variation) 🧵[5/N]

..., as verified in our novel view synthesis experiments. 🧵[4/N]

In contrast, Robust e-NeRF is able to directly and robustly reconstruct NeRFs under various real-world conditions, especially from sparse and noisy events generated under non-uniform motion, ... 🧵[3/N]

Prior work on NeRF reconstruction from a moving event camera is mainly limited in terms of: 1. Dependence on dense and low-noise event streams 2. Generalization to arbitrary contrast threshold values and camera speed profiles 🧵[2/N]

Super excited to share my #ICCV2023 work with Gim Hee Lee (@gimhee_lee) on "Robust e-NeRF: NeRF from Sparse & Noisy Events under Non-Uniform Motion"! Project page: https://t.co/YP35o4ASjn Paper: https://t.co/PLtqliQH1q Code: https://t.co/bGCsYWSCBV 🧵[1/N]

Our code can be found at https://t.co/tz7aZj9RhN. We will also be presenting this work at #ECCV2022 (3:30PM - 5:30PM, 27 Oct @ Poster #64, Hall B). Drop by to find out more! 🧵[4/N=4]

In contrast, prior works generally predefine the parametric domain, which unnecessarily constrains the boundary and topology of each chart 🧵[3/N]

This is enabled by a fully learnable parametric domain, given by a probabilistic occupancy field defined on the (-1, 1)² open square of the UV parametric space 🧵[2/N]

Thanks @neural_fields for featuring our work! Minimal Neural Atlas is a novel explicit neural surface representation that can effectively learn a minimal atlas of 3 charts with distortion-minimal parameterization for surfaces of complex topology. 🧵[1/N]