Lightnews — Scholar-powered news

Oisin Mac Aodha @oisinmacaodha.bsky.social · 19d

We have some fantastic invited speakers lined up:
sites.google.com/g.harvard.ed...

In addition, we have a paper track for both novel unpublished and published work (deadline Oct 10th):
sites.google.com/g.harvard.ed...

3

Oisin Mac Aodha @oisinmacaodha.bsky.social · 19d

Working at the intersection of AI and Climate/Conservation?

If so, check out our upcoming workshop which will taking place at #EurIPS in Copenhagen in December 2025.
sites.google.com/g.harvard.ed...

@euripsconf.bsky.social
‬

Climate AI Nordics @climateainordics.com · 19d

🌍 Excited to announce our Workshop on AI for Climate & Conservation (AICC) at #EurIPS2025 in Copenhagen! 🎉

📢 Call for Participation: sites.google.com/g.harvard.ed...

Confirmed speakers from Mistral AI, DeepMind, ETH Zurich, LSCE & more.

Looking forward to meeting and discussing in Copenhagen!

1 3 10

Oisin Mac Aodha @oisinmacaodha.bsky.social · 19d

Very cool!

1

Oisin Mac Aodha @oisinmacaodha.bsky.social · 27d

The US-UK Scientific Forum on Measuring Biodiversity for Addressing the Global Biodiversity Crisis summary report is now available.

I participated earlier this year and it was really enlightening. The report covers explores tools, challenges, and solutions for tackling the biodiversity crisis.

National Academy of Sciences @nasonline.org · 27d

How do we measure life on Earth? 🌍🌱 The new summary report from the 2025 US-UK Forum on Measuring Biodiversity, hosted by the NAS and @royalsociety.org, explores tools, challenges, and solutions for tackling the #biodiversity crisis.

Read here: bit.ly/460RB7R

3

Oisin Mac Aodha @oisinmacaodha.bsky.social · Jul 18

Check out the paper for more details:

Feedforward Few-shot Species Range Estimation
arxiv.org/abs/2502.14977
Lange et al. ICML 2025

Oisin Mac Aodha @oisinmacaodha.bsky.social · Jul 18

This work was led by Christian Lange, with support from our amazing collaborators Max Hamilton, Elijah Cole, Alexander Shepard, Samuel Heinrich, Angela Zhu, Subhransu Maji, Grant Van Horn, and Oisin Mac Aodha.

1

Oisin Mac Aodha @oisinmacaodha.bsky.social · Jul 18

FS-SINR is efficient. At test time, it can take an arbitrary number of observations (i.e., context locations) as input, along with optional metadata, and generate a predicted range in a single forward pass of the model.

1

Oisin Mac Aodha @oisinmacaodha.bsky.social · Jul 18

We obtain better performance in the few-shot setting, i.e., where we have very limited observations for a species. In the x-axis in this plot we vary the number of observations provided to each model for a set of different species and on the y-axis we measure the quality of the range predictions.

1

Oisin Mac Aodha @oisinmacaodha.bsky.social · Jul 18

We observe improved range prediction performance compared to existing methods, e.g., SINR from Cole et al. at ICML 2023 or LE-SINR from Hamilton et al. at NeurIPS 2024.

Top row: Gabar Goshawk
Bottom row: Black-naped Monarch

1

Oisin Mac Aodha @oisinmacaodha.bsky.social · Jul 18

In this example, we see a prediction for FS-SINR using a single presence observation as input shown as a white dot (left). Conditioning the model with text (e.g. middle and right), can dramatically change the range predictions.

1

Oisin Mac Aodha @oisinmacaodha.bsky.social · Jul 18

FS-SINR can be conditioned on in-situ presence observations for a species not seen during training in addition to text descriptions of their ranges or images of the species if available.

1

Oisin Mac Aodha @oisinmacaodha.bsky.social · Jul 18

In the previous video, we illustrated test time range predictions for FS-SINR for the European Robin where we vary the number of presence observations (shown as white circles). As more observations are added, the predictions improve, becoming more similar to the expert range map (right bottom).

1

Oisin Mac Aodha @oisinmacaodha.bsky.social · Jul 18

This week at #ICML we are presenting our new work titled Feedforward Few-shot Species Range Estimation.

TLDR;
* Our model, FS-SINR, can estimate a species' range from few observations
* It does not require an retraining for previously unseen species
* It can integrate text and image information

1 2 7

Reposted by Oisin Mac Aodha

FGVC Workshop @fgvcworkshop.bsky.social · Jul 17

We are not missing out! FGVC12 is excited to support "FOMO25: Foundation Model Challenge for Brain MRI".
This MICCAI25 challenge is still running and there is still time to participate!

Submission deadline: August 20, 2025
Join here: fomo25.github.io
Check out the thread below👇

2 4

Oisin Mac Aodha @oisinmacaodha.bsky.social · Jun 15

Paper:
arxiv.org/abs/2412.04120

Webpage:
iamsalvatore.io/cross_sdf/

1 2

Oisin Mac Aodha @oisinmacaodha.bsky.social · Jun 15

CrossSDF: 3D Reconstruction of Thin Structures From Cross-Sections

We will be presenting our work on thin structure reconstruction at the final poster session (4-6pm) at #CVPR2025 today.

Stop by poster #457 to learn more.

1 4 20

Oisin Mac Aodha @oisinmacaodha.bsky.social · Jun 14

Paper:
arxiv.org/abs/2411.17385

Project page:
danier97.github.io/depthcues/

Work led by Duolikun Danier:
danier97.github.io

Oisin Mac Aodha @oisinmacaodha.bsky.social · Jun 14

To explore this question, we developed a new benchmark, DepthCues, to evaluate human-like monocular depth cues in large vision models. We show that strong performance on these cues emerges in more recent larger monocular depth models.

1

Oisin Mac Aodha @oisinmacaodha.bsky.social · Jun 14

DepthCues: Evaluating Monocular Depth Perception in Large Vision Models

Do automated monocular depth estimation methods use similar visual cues to humans?

To learn more, stop by poster #405 in the evening session (17:00 to 19:00) today at #CVPR2025.

1 1 5

Oisin Mac Aodha @oisinmacaodha.bsky.social · Jun 14

Work led by @sizquierdo.bsky.social

1

Oisin Mac Aodha @oisinmacaodha.bsky.social · Jun 14

Paper:
arxiv.org/abs/2503.22430

Code:
github.com/nianticlabs/...

1 1

Oisin Mac Aodha @oisinmacaodha.bsky.social · Jun 14

MVSAnywhere: Zero-Shot Multi-View Stereo

Looking for a multi-view stereo depth estimation model which works anywhere, in any scene, with any range of depths?

If so, stop by our poster #81 today in the morning session (10:30 to 12:20) at #CVPR2025.

1 2

Oisin Mac Aodha @oisinmacaodha.bsky.social · Jun 12

Check out Nikolas @tsagkas.bsky.social and Danier's work on understanding the limitations of pre-trained visual representations for visuomotor robot learning at the Embodied AI Workshop at CVPR 2025 today.

Nikolas Tsagkas @tsagkas.bsky.social · Jun 12

Danier Duolikun presents our work on pre-trained visual representations for visuomotor robot learning today at #CVPR2025 in the 6th Embodied AI Workshop!
🗣️ Talk: 15:30, Room 101 D
📌 Poster: 12:00–13:30, ExHall D (#140–169)
Come say hi!

More info here: tsagkas.github.io/pvrobo/

4

Oisin Mac Aodha @oisinmacaodha.bsky.social · Jun 11

You can find Room 104E on Level 1 (i.e. street level).

Oisin Mac Aodha @oisinmacaodha.bsky.social · Jun 11

We have a great line up of fantastic speakers.

1