←
Jerred Chen26 posts
Jerred Chen avatar

Jerred Chen

@JerredChen

PhD student at @UniofOxford

◷ Joined February 2021▧ 8 Photos and videos
Ronnie Clark avatar
⟳ Jerred Chen retweeted
Ronnie Clark@ronnieclark__·May 28

Introducing 𝗣𝗜𝗫𝗟𝗥𝗲𝗹𝗶𝗴𝗵𝘁 - our new method for precise image relighting. Our method allows you to specify the lighting you want exactly as in a standard 3D pipeline (no text prompts or exemplars). Try the interactive demo here: mlfarinha.github.io/pixl-rel… w/Miguel Farinha

Jacob Lin avatar
⟳ Jerred Chen retweeted
Jacob Lin@jacob_lin2·27 Nov 2025

Interested in how 3D vision can help the development of high-resolution weather/climate models? In our NeurIPS 2025 spotlight paper, we show how ground-based cameras can be used to estimate physically-meaningful 3D cloud properties at a high spatiotemporal resolution. 🧵 1/7

0:07
Jerred Chen avatar
Jerred Chen@JerredChen·28 Jun 2025

Now accepted in ICCV2025! See you in Hawaii 🌺

Motion blur typically breaks SLAM/SfM algorithms - but what if blur was actually the key to super-robust motion estimation? In our new work, Image as an IMU, @ronnieclark__ and I demonstrate exactly how a single motion-blurred image can be used to our advantage. 🧵1/9

0:16
Jerred Chen avatar
Jerred Chen@JerredChen·26 Mar 2025

Motion blur typically breaks SLAM/SfM algorithms - but what if blur was actually the key to super-robust motion estimation? In our new work, Image as an IMU, @ronnieclark__ and I demonstrate exactly how a single motion-blurred image can be used to our advantage. 🧵1/9

0:16
Jerred Chen avatar
Jerred Chen@JerredChen·26 Mar 2025

This project wouldn't have been possible without my supervisor @ronnieclark__ whose help and guidance I'm very grateful for, and thanks also to Linacre College and the CS Department for allowing us to use their spaces to record our test and training scenes. 9/9

Jerred Chen avatar
Jerred Chen@JerredChen·26 Mar 2025

While baselines (e.g. COLMAP, MASt3R) require two or more images to estimate camera motion, we are much more accurate and robust even when our network only uses a *single image as input*! For example, see how our method compares to MASt3R with pairwise views. 7/9

0:07
Jerred Chen avatar
Jerred Chen@JerredChen·26 Mar 2025

We evaluate on 4 unseen, in-the-wild videos and use the Strayscanner app to obtain videos and GT. Since estimating the true velocity direction from 1 img is ambiguous, we flip the direction if necessary based on the photometric error between the current and prev/next frames. 6/9

0:04
Jerred Chen avatar
Jerred Chen@JerredChen·26 Mar 2025

To train our method, we use a subset of ScanNet++v2 to synthesize motion blurred imgs with GT poses, depths and flow. Because our model is fully differentiable, we also collect real-world imgs to further finetune our model end-to-end, supervised only on ARKit poses. 5/9

Post attachment
Jerred Chen avatar
Jerred Chen@JerredChen·26 Mar 2025

Our method is simple: we train a network that outputs flow and depths, which we use with the intrinsics to form a linear system under the small motion assumption. We solve for relative pose using LLS in a fully differentiable manner and use the exposure to compute velocity. 4/9

Post attachment
Jerred Chen avatar
Jerred Chen@JerredChen·26 Mar 2025

Our key insight: the motion blur pattern encodes the trajectory of each 3D point - creating implicit optical flow or “virtual correspondences”. Extracting these provides a powerful cue for robust camera motion estimation! 3/9

Post attachment
Jerred Chen avatar
Jerred Chen@JerredChen·26 Mar 2025

Motion blur occurs when the camera or scene moves during exposure, washing out feature detail. This is problematic for standard pose estimation methods, which assume instantaneous, sharp exposures. 2/9

0:15
Jerred Chen avatar
Jerred Chen@JerredChen·26 Mar 2025

Motion blur typically breaks SLAM/SfM algorithms - but what if blur was actually the key to super-robust motion estimation? In our new work, Image as an IMU, @ronnieclark__ and I demonstrate exactly how a single motion-blurred image can be used to our advantage. 🧵1/9

0:16