Paul Henderson

Associate Prof. in Machine Learning at University of Glasgow. I research computer vision/graphics, deep generative models, ML for sciences & healthcare

[Nat Friedman @natfriedman]

Vesuvius Challenge has been quietly making a lot of progress; we read a full scroll, developed a new unwrapping algorithm with a global spiral prior, and we just launched a new $1M grand prize for reading one of the scrolls we've scanned.

There are also 10 $50k prizes available for finding text in a scroll we haven't found text in before, plus monthly progress prizes.

Join the discord!

[Vesuvius Challenge @scrollprize]

A new phase BEGINS! 🚀

🏆 $1M Grand Prize
💰 $2M+ in open prizes
👑 $20K awarded every month to the best submission
https://t.co/P6VuhVYK1L

Will you read the next scroll? 👀
Spread the word and repost! 🔁

[Rod Murray-Smith @MurraySmithRod]

There are two new permanent posts available at @GlasgowCS in AI/ML especially with HCI/ML in Science/Active Inference focus. Deadline 17th Feb
https://t.co/m30arKeDIc

[Paul Henderson @pmh47_ml]

@zhuhu_polyu Yes, soon after it's accepted for publication

[Paul Henderson @pmh47_ml]

New preprint! TL;DR = latent diffusion over 3d gaussian splats enables fast scene generation (0.2s for 50 ddim steps) from 0/1/few images

https://t.co/SySb0xPiux

[British Machine Vision Conference (BMVC) @BMVCconf]

4 months to go for #BMVC2024! 🥳This is a sneak peak of the beautiful #Glasgow city chambers where we will be hosting the drink reception ! 🍻#VisitGlasgow Stay tuned for more updates!

[BMVA Summer School @BmvaCvss]

Dr. Paul Henderson @pmh47_ml delivered a talk on "Structured Generative Models for Computer Vision" in the 27th BMVA Summer School in Durham University.

[MrNeRF @janusch_patas]

Sampling 3D Gaussian Scenes in Seconds with Latent Diffusion Models
https://t.co/FrFzRM8HKu

[Paul Henderson @pmh47_ml]

Joint work with @anciukevicius at @InfAtEd, and @GlasgowCS students Melonie de Almeida & Daniela Ivanova

@IDAglasgow @CVAS_UofG

[Paul Henderson @pmh47_ml]

We design an autoencoder that maps multi-view images to 3D Gaussian splats, and simultaneously builds a compressed latent representation of these splats. Then, we train a multi-view diffusion model over the latent space to learn an efficient generative model

[Paul Henderson @pmh47_ml]

Trained end-to-end with no segmentation masks, and no pretrained depth predictors

Project: https://t.co/BhnlZz8RBS

arXiv: https://t.co/37yOLqf3x4

[Paul Henderson @pmh47_ml]

In Vienna? Interested in 3D generative models? Come chat at @anciukevicius's @iclr_conf #ICLR2024 poster later (Tuesday 16:30, #42) on Denoising Diffusion via Image-based Rendering
https://t.co/gVVmCv9rlf

[Paul Henderson @pmh47_ml]

Happy to announce our work on generative 3D reconstruction "Denoising Diffusion via Image-Based Rendering" is accepted at @iclr_conf.

Created by @Anciukevicius, partly during his Google-internship with @fedassa and Fabian Manhardt

A quick summary...

@CVAS_UofG @IDAglasgow

[Paul Henderson @pmh47_ml]

We show how to learn 3D diffusion models over large real-world scenes using only multiview images for training. It's standard diffusion training, BUT with a twist - the denoiser NN predicts a latent 3D scene representation before rendering to give clean pixels

[Paul Henderson @pmh47_ml]

It's similar in flavour to Viewset Diffusion (nice concurrent work from @StanSzymanowicz & @chrirupp) -- but ours avoids the bounded 3D feature volume. The image-based reprn enables generating 3D content anywhere that any camera sees, hence don't need to mask away background

[Paul Henderson @pmh47_ml]

Happy to announce our work on generative 3D reconstruction "Denoising Diffusion via Image-Based Rendering" is accepted at @iclr_conf.

Created by @Anciukevicius, partly during his Google-internship with @fedassa and Fabian Manhardt

A quick summary...

@CVAS_UofG @IDAglasgow

[Paul Henderson @pmh47_ml]

@iclr_conf @CVAS_UofG @InfAtEd @Anciukevicius See the project page https://t.co/BhnlZz8RBS for more examples rendered at much higher resolution (1024x1024 -- higher than any similar model)

[Paul Henderson @pmh47_ml]

...and here are the results on generative reconstruction from *one* input image, which I forgot before =)
This was trained end-to-end from images only, yet learns to create plausible details even in occluded areas

@iclr_conf @CVAS_UofG @InfAtEd @Anciukevicius

[Paul Henderson @pmh47_ml]

Happy to announce our work on generative 3D reconstruction "Denoising Diffusion via Image-Based Rendering" is accepted at @iclr_conf.

Created by @Anciukevicius, partly during his Google-internship with @fedassa and Fabian Manhardt

A quick summary...

@CVAS_UofG @IDAglasgow

[Paul Henderson @pmh47_ml]

Project: https://t.co/BhnlZz8RBS

arXiv: https://t.co/37yOLqf3x4

@GlasgowCS #ICLR2024

[Paul Henderson @pmh47_ml]

Happy to announce our work on generative 3D reconstruction "Denoising Diffusion via Image-Based Rendering" is accepted at @iclr_conf.

Created by @Anciukevicius, partly during his Google-internship with @fedassa and Fabian Manhardt

A quick summary...

@CVAS_UofG @IDAglasgow

[Paul Henderson @pmh47_ml]

In this #ICLR2024 work, we use an image-based scene representation we call IB-planes, where features unprojected from each camera are fused to by an MLP to define local appearance/density (similar to PixelNeRF)