Jack (in SF) Langerman

Very proud that our paper has been accepted to
@iccv.bsky.social !!!

See you in Hawaii!
bsky.app/profile/duch...

🎉 Thrilled to share our CVPR 2025 Award Candidate & Oral paper:

🔹 GlobustVP
Convex Relaxation for Robust Vanishing Point Estimation in Manhattan World

🧱 Global optimality
💥 Tolerates up to 70% outliers
⚡ Fast runtime

📄 Paper: arxiv.org/abs/2505.04788

💻 Code: github.com/WU-CVGL/GlobustVP

1/

#ICCV2025 reviews are out and being sent via email to authors! They will also be available on OpenReview later. 11,152 active submissions all have at least 3 reviews. Authors have the opportunity to submit a rebuttal by May 16 2025 11:59 PM HST.

Please reshare, tell you friends, give it a try, and don't hesitate to reach out to the team with any questions!

bsky.app/profile/jack...

🌍 Workshop + other challenges at:
usm3d.github.io
Let’s make SfM more structured.
Good luck, teams!

(7/7)

usm3d/handcrafted_submission_2025 · Hugging Face We’re on a journey to advance and democratize artificial intelligence through open source and open science.

📌 Want a clear reference?
Check out the example baseline submission:
huggingface.co/usm3d/handcr...

(6/7)

Explaining Human Preferences via Metrics for Structured 3D Reconstruction "What cannot be measured cannot be improved" while likely never uttered by Lord Kelvin, summarizes effectively the purpose of this work. This paper presents a detailed evaluation of automated metrics ...

📊 NEW Evaluation Metric
Say goodbye to WED.
Hello to a human-aligned combo of:
• Vertex F1
• Edge IoU
Backed by this paper:

👉 arxiv.org/abs/2503.08208

GitHub - s23dr/hoho2025: Tools and utilities for the S23DR-2025 competition and HoHo25k Dataset Tools and utilities for the S23DR-2025 competition and HoHo25k Dataset - s23dr/hoho2025

⚙️ All helper functions + data loaders:
pip install hoho2025
🛠️ Codebase: github.com/s23dr/hoho2025

(4/7)

usm3d/hoho25k · Datasets at Hugging Face We’re on a journey to advance and democratize artificial intelligence through open source and open science.

📦 New dataset:
hoho25k — 5x bigger than 2024!
• 25,000 scenes
• 8 images per scene
• 200,000 total images
📥 Get it here: huggingface.co/datasets/usm...

(3/7)

S23DR2025 - a Hugging Face Space by usm3d This application allows you to view competition details, manage your submissions, and check leaderboards. You can access rules, dataset information, and view your own submissions and scores.

🧠 The task:
Turn multiview inputs (semantic segs, monocular depth) + SfM outputs (point clouds + camera poses) into sparse, geometrically accurate wireframes.
AKA: “More Structured Structure-from-Motion” 😎

Challenge page 👉 huggingface.co/spaces/usm3d...

(2/7)

🚨 Just one month left to submit your solutions for The Structured Semantic 3D Reconstruction (S23DR-2025) Challenge!!! It is not too late to join!

Comp on @hf.co, part of the Workshop on Urban Scene Modeling at @cvprconference.bsky.social 2025

🔥$25,000 prize pool. Deadline: June 5, 2025.

🧵 (1/7)

LoRA+: Efficient Low Rank Adaptation of Large Models Soufiane Hayou, Nikhil Ghosh, Bin Yu In this paper, we show that Low Rank Adaptation (LoRA) as originally introduced in Hu et al. (2021) leads to suboptimal finetuning of models with large width (embedding dimension). This is due to the fact that adapter matrices A and B in LoRA are updated with the same learning rate. Using scaling arguments for large width networks, we demonstrate that using the same learning rate for A and B does not allow efficient feature learning. We then show that this suboptimality of LoRA can be corrected simply by setting different learning rates for the LoRA adapter matrices A and B with a well-chosen ratio. We call this proposed algorithm LoRA+. In our extensive experiments, LoRA+ improves performance (1-2 % improvements) and finetuning speed (up to ∼ 2X SpeedUp), at the same computational cost as LoRA.

hey @cloneofsimo.bsky.social (or anyone else) did you ever try LoRA+ style LR split between down/up LoRA projection matrices on diffusion models?

Logo for MIB: A Mechanistic Interpretability Benchmark

Lots of progress in mech interp (MI) lately! But how can we measure when new mech interp methods yield real improvements over prior work?

We propose 😎 𝗠𝗜𝗕: a 𝗠echanistic 𝗜nterpretability 𝗕enchmark!

When I was working at Bell Labs in Murray Hill we used to specifically go to 3rd floor for coffee because the couch in that particular coffee room was (supposedly) Dennis Ritchie's.

was also a room with a whiteboard that may have been where Shannon did a lot of the info theory work.

good times

do you think this is a matter of semantic disipline or genuine beliefs about qualia?

ie "with certain inputs models can produce outputs that contain patterns typically associated with anxiety in humans. asking them to perform mindfulness exercises can mitigate this behavior" vs "they feel anxious"?

1 week to USM3D deadline

more blending between rings i think (~long range dependence)

Explaining Human Preferences via Metrics for Structured 3D Reconstruction

@jacklangerman.bsky.social Denys Rozumnyi, Yuzhong Huang, @ducha-aiki.bsky.social

tl;dr: we asked 3D modelers to rank wireframe reconstructions & compared it to ranking by metrics. Observations🧵
1/
arxiv.org/abs/2503.08208

2nd Building3D #CVPR2025 challenge at #USM3D workshop is open!
Task: point cloud to wireframe.
Prize pool: $10k
Competition deadline: May 25 2025.
Website: huggingface.co/spaces/Build...
@cvprconference.bsky.social

We made a new keypoint detector named DaD, paper isn't up yet, but code and weights are:
github.com/Parskatt/dad

Those, who work in structured (images/pcl to CAD) reconstruction - USM3D #CVPR2025 workshop submissions are open.
Deadline: March 24 2025
Both full papers (8 pages) and extended abstracts (4 pages) are OK

usm3d.github.io
#USM3D
@cvprconference.bsky.social @jacklangerman.bsky.social

We have extended the deadline for paper submission for Image Matching Workshop. Anything image matching or 3D reconstruction related is welcomed
Now it is March 17
@cvprconference.bsky.social
#CVPR2025
image-matching-workshop.github.io

MPS operator coverage tracking issue (2.6+ version) · Issue #141287 · pytorch/pytorch 🐛 Describe the bug This issue is to have a centralized place to list and track work on adding support to new ops for the MPS backend. PyTorch MPS Ops Project : Project to track all the ops for MPS ...

github.com/pytorch/pyto...

yeah `PYTORCH_ENABLE_MPS_FALLBACK=1 python myscript.py` did it.

bsky.app/profile/jack...

:_( operation not implemented on MPS woe is me...

oh no....

what is the best intro I can send people who suddenly want to know "what is AI?" and "what is a transformer based model?" but arent really going to invest tons of time and maybe dont have any math background?

my default is 3b1b, but does anyone have another suggestion?

Image matching and ChatGPT - new post in the wide baseline stereo blog.

tl;dr: it is good, even feels like human, but not perfect.
ducha-aiki.github.io/wide-baselin...

ive said for a long time this is one of the things I love most about ML: get to play in everyones' playgrounds :-)

o u know - jc