harpreet

Hacker-in-residence @voxel51. open source contributor. shipping fiftyone integrations. cvpr is better than neurips

[Jimmy at Voxel51 @jimmy_voxel51]

EvoLogics uses Voxel51’s FiftyOne Enterprise to prepare the right training data and evaluate perception models for autonomous subsea missions, giving the team greater control and speed from data to production.

Learn more about how FiftyOne can optimize your computer vision workflows by booking a demo: https://t.co/IUXbrtea3I

Learn more about Evologics: https://t.co/mrvFGixqTV

EvoLogics designs and manufactures underwater communication systems, positioning networks, and autonomous vehicles for the world's most demanding subsea environments. Based in Germany, the company has built technology for clients across commercial, offshore, defense, and research sectors.

Operating at depths and in conditions unreachable by conventional means, EvoLogics systems locate survivors in search and rescue operations, inspect subsea pipelines and infrastructure, monitor ocean environments, and support naval mine countermeasure activities. Their systems process multimodal data, including sonar data from various types, underwater and surface-mounted cameras, LiDARs, and hydrophones, in real time, during live missions.

#computervision #ai #artificialintelligence #machinevision
#machinelearning #datascience #physicalai #mcp #agents

[harpreet @DataScienceHarp]

@SanSanychUA Hi, yes it is already on the hf repo actually I added the code there.

[harpreet @DataScienceHarp]

great work from @CodyJzr

3d point cloud reconstruction wasn't part of the original Syn4D dataset, but it was possible to reconstruct it from the ground-truth annotations that were included:

> Read per-frame depth (float32 EXR), RGB images, and per-frame camera intrinsics + extrinsics (focal length, sensor size, position, yaw/pitch/roll) from all 8 synchronised camera views

> Applied sRGB gamma correction to the linear-space RGB renders so colours display correctly

> Back-projected each valid depth pixel into a shared Unreal Engine world coordinate system using the standard pinhole camera model, converting the result from centimetres to metres

> Coloured each 3D point from its corresponding RGB pixel, merged all 8 views, then voxel-downsampled and removed statistical outliers to produce a clean cloud per sequence

[harpreet @DataScienceHarp]

static 3D reconstruction is mostly solved. dynamic scenes, where objects move and people walk around, that's still an open problem.

the bottleneck is data: you need multiple synchronized cameras capturing the same moment from different angles with dense ground truth

Syn4D is a fully synthetic multiview dataset built for this. 8 synchronized cameras, Unreal Engine 5, per-frame depth maps, instance segmentation, camera poses, and natural language captions across offices, warehouses, and hospitals

i grouped the 8 camera views together in fiftyone with 3D point cloud reconstructions so you can flip between any camera angle, the depth and segmentation overlays, and the fused 3D scene for any sequence

check out the dataset here: https://t.co/0tmVGflYQR

btw if you're at ICRA next week hmu or come by booth or swing by booth B081 and say hi

#ICRA2026

[harpreet @DataScienceHarp]

the 3d reconstruction wasn't part of the original dataset but i was able to reconstruct it by:

> reading per-frame exr depth, rgb images, and camera poses from each scene
>back-project depth pixels into a shared world coordinate system
>color each 3d point from the RGB image and merge frames into one cloud

check out the dataset here: https://t.co/X8QoGjBwrG

if you're at icra this year hmu, or swing by booth B081 and say hi

[harpreet @DataScienceHarp]

robots can't grasp transparent objects because depth sensors can't see them

glass, bottles, clear containers just disappear

ClearDepth is a stereo dataset built for this problem: left/right video pairs with ground truth depth, surface normals, instance segmentation, and camera poses across 204 indoor scenes

i built reconstructed point cloud reconstructions for each scene and grouped everything in fiftyone so you can flip between the stereo views, the dense labels, and the 3D reconstruction side by side

#ICRA2026

[Jimmy at Voxel51 @jimmy_voxel51]

Level up your computer vision workflows with a free hands-on workshop for your team! Book a workshop: https://t.co/hLB5psNaDN

These hands-on workshops are delivered by Voxel51 computer vision experts. Both virtual and in-person formats.

* 60 min virtual workshop

* Half-day onsite workshop

* Full-day onsite workshop and hackathon

#mcp #skills #computervision #ai #artificialintelligence #machinevision #machinelearning #physicalai

[harpreet @DataScienceHarp]

cc @xshocng1

[harpreet @DataScienceHarp]

what a 3D reconstruction of a transparent object scene looks like when you back-project 121 frames of diffusion-estimated depth into a point cloud

built from TransPhy3D — each sequence has RGB video + depth + normals + camera calibration, all grouped in FiftyOne

#ICRA2026

[harpreet @DataScienceHarp]

check out the dataset here: https://t.co/toApD7jyty

hmu if you're gonna be at icra, or swing by booth B081 and say hi

[harpreet @DataScienceHarp]

@PINTO03091 Awesome thank you for your contributions. I have only become aware of your work today, cheers!

[harpreet @DataScienceHarp]

this came out of a workshop I ran at an enterprise. enough teams have been asking me to come train their people on fiftyone that we made it a program.

curation, annotation, evaluation, debugging, whatever you're working on: https://t.co/EQOe80djb4

[harpreet @DataScienceHarp]

inding the right images to annotate in a giant unlabeled dataset shouldn't take a week

built a fiftyone panel that does it from a handful of reference crops. ranked dataset, heatmaps, tagged queue.

https://t.co/WbAFvDZWpF

[harpreet @DataScienceHarp]

Agree

https://t.co/onQWGvfE67

[Jiawei Yang @JiaweiYang118]

Deserve another best paper

[harpreet @DataScienceHarp]

great work from @jianyuan_wang @n_karaev @davnov134 et. al.

you can quickly get started with the model in fiftyone by following the steps in this repo: https://t.co/rPqYDOPUoq

[harpreet @DataScienceHarp]

wow, vggt-omega is dope. might just take best paper at cvpr again

[Jimmy at Voxel51 @jimmy_voxel51]

Missed the April 30 - Best of WACV Meetup? Then you missed Brent Griffin introducing "Zcore," a state-of-the-art core set selection approach designed to optimize data set usage. Learn more: ​​https://t.co/gYVdDFygaK

The primary goal of core set selection is to potentially eliminate 90% of a data set before spending time on labeling or model training.

#mcp #skills #computervision #ai #artificialintelligence #machinevision #machinelearning #physicalai

[Jimmy at Voxel51 @jimmy_voxel51]

Join us on May 12 for day two of the Best of 3DV 2026 series of virtual events.

Register for the Zoom: https://t.co/GTcSAw9t0j

Talks will include:

* Precise lighting control in diffusion models by drawing shadows - Frédéric Fortier-Chouinard at Laval University

* SmokeSeer: 3D Gaussian Splatting for Smoke Removal and Scene Reconstruction - Neham Jain at Meshy AI

* Online Video Depth Anything: Temporally-Consistent Depth Prediction with Low Memory Consumption - Johann-Friedrich Feiden at Heidelberg University

* Broadening View Synthesis of Dynamic Scenes from Constrained Monocular Videos - Le Jiang at Northeastern University AClab

***********

Want to build better computer vision models? FiftyOne is an open source toolkit from Voxel51 (our Meetup sponsor) that helps you curate datasets, evaluate model performance, visualize embeddings, catch annotation errors, and eliminate duplicate images—all in one place.

“pip install fiftyone” is all it takes to get started - https://t.co/am1W0dM5Lc

#mcp #skills #computervision #ai #artificialintelligence #machinevision #machinelearning #physicalai

[Jason Corso @_JasonCorso_]

Show me the impact not the paper count. Paper mills outputting LPUs are a cancer to the field.

And the data seems incomplete. In aggregate what about all of the institutions less than paper threshold here.