Also, our follow-up work and extension to 3D point-clouds has been presented at CVPR2025:
openaccess.thecvf.com/content/CVPR...

This work has seen many R2's and the in meantime other methods have appeared with the name "vocabulary-free". While the input is indeed vocabulary-free, the method is not, it has to be auto-generated in order to enumerate all objects in a scene.

Learn more today @ #ICCV2025:
📅 #418 Poster Session 5 10:45-12:45 HST
📍 Exhibit Hall I
🌐 ozzyou.github.io/autoseg.gith...

We present "Auto-vocabulary segmentation" in which the vocabulary is auto-generated with a captioning method which is then used by the segmentation method.

We also introduce an automatic vocabulary mapping for evaluation on human-labeled datasets which typically have a much smaller vocabulary.

A few more impression from the Neural SLAM workshop today.
Thanks so much to our keynote speakers for making this such an insightful and memorable event:
- Luca Carlone @lucacarlone.bsky.social
- Angjoo Kanazawa @akanazawa.bsky.social
- Federico Tombari @
- Jakob Engel @jajuengel
👏👏🙏

Our fourth keynote today by Jakob Engel.
📍 Room 304 A
🌐 sites.google.com/view/neuslam...
#ICCV2025

Federico Tombari during his keynote right now.
📍 Room 304 A
🌐 sites.google.com/view/neuslam...
#iccv25

Angjoo Kanazawa giving her keynote talk now.
📍 Room 304 A
🌐 sites.google.com/view/neuslam...