(v) INSID3 outperforms both training free methods relying on SAM, and fine-tuned approaches, which fall short out-of-domain, demonstrating strong generalization across tasks and domains.

(iv) INSID3 removes it in a training-free way: we identify the positional component of DINOv3 features and project onto its orthogonal complement.

(iii) Besides semantic matches, DINOv3 also responds to absolute image position. Given a patch on the bird’s tail in the reference image, the DINOv3 similarity map activates on (i) the tail in the target image, but also (ii) over the left portion of the image.

(ii) Dense DINOv3 features naturally induce a structured decomposition of the scene. By clustering them, we obtain coherent object- and part-level regions without supervision.

🔥 Can in-context segmentation emerge directly from frozen DINOv3 features?

At #CVPR2026, we present INSID3: Training-Free In-Context Segmentation with DINOv3 — a collaboration between PoliTo, TU Darmstadt and TU Munich.

Check it out: github.com/visinf/INSID3

📄 Read the full paper:
SANSA: Unleashing the Hidden Semantics in SAM2 for Few-Shot Segmentation
Now on arXiv → arxiv.org/abs/2505.21795

🔹 4/4 – Promptable segmentation in action SANSA reduces reliance on costly pixel-level masks by supporting point, box, and scribble prompts
📈enabling fast, scalable annotation with minimal supervision.
See the qualitative results 👇

🔹 3/4 – SANSA achieves state-of-the-art in few-shot segmentation. We outperform specialist and foundation-based methods across various benchmarks:
📈 +9.3% mIoU on LVIS-92i
⚡ 3× faster than prior works
💡 Only 234M parameters (4-5x smaller than competitors)

🔹2/4 – Unlocking semantic structure

SAM2 features are rich, but optimized for tracking.
🧠 Insert bottleneck adapters into frozen SAM2
📉 These restructure feature space to disentangle semantics
📈 Result: features cluster semantically—even for unseen classes (see PCA👇)

🚀 As #CVPR2025 week kicks off, meet SANSA: Semantically AligNed Segment Anything 2
We turn SAM2 into a semantic few-shot segmenter:
🧠 Unlocks latent semantics in frozen SAM2
✏️ Supports any prompt: fast and scalable annotation
📦 No extra encoders

📎 github.com/ClaudiaCutta...

#ICCV2025

I guess merging the events could also work 😂 I wonder whether cricket players would be better at ComputerVision than CV researchers are at cricket, or viceversa

To Match or Not to Match: Revisiting Image Matching for Reliable Visual Place Recognition

Davide Sferrazza, @berton-gabri.bsky.social Gabriele Trivigno, Carlo Masone
tl;dr: global descriptors nowadays are often better than local feature matching methods for simple datasets.
arxiv.org/abs/2504.06116

✨ SAMWISE achieves state-of-the-art performance across multiple #RVOS benchmarks—while being the smallest model in RVOS! 🎯 It also sets a new #SOTA in image-level referring #segmentation. With only 4.9M trainable parameters, it runs #online and requires no fine-tuning of SAM2 🚀

🚀 Contributions:
🔹 Textual Prompts for SAM2: Early fusion of visual-text cues via a novel adapter
🔹 Temporal Modeling: Essential for video understanding, beyond frame-by-frame object tracking
🔹 Tracking Bias: Correcting tracking bias in SAM2 for text-aligned object discovery

🔥 Our paper SAMWISE: Infusing Wisdom in SAM2 for Text-Driven Video Segmentation is accepted as a #Highlight at #CVPR2025! 🎉
We make #SegmentAnything wiser, enabling it to understand textual prompts—training only 4.9M parameters! 🧠
💻 Code, models & demo: github.com/ClaudiaCutta...

Why SAMWISE?👇

To Match or Not to Match: Revisiting Image Matching for Reliable Visual Place Recognition

Davide Sferrazza, @berton-gabri.bsky.social, @gabtriv.bsky.social, Carlo Masone

tl;dr:VPR datasets saturate;re-ranking not good;image matching->uncertainty->inlier counts->confidence

arxiv.org/abs/2504.06116

🚀 Paper Release! 🚀
Curious about image retrieval and contrastive learning? We present:

📄 "All You Need to Know About Training Image Retrieval Models"
🔍 The most comprehensive retrieval benchmark—thousands of experiments across 4 datasets, dozens of losses, batch sizes, LRs, data labeling, and more!

Trying to convince my bluesky feed to put me in the Computer Vision community. Right now I only see posts about the orange-haired president. @berton-gabri.bsky.social @gabrigole.bsky.social how did you do it?

Image segmentation doesn’t have to be rocket science. 🚀
Why build a rocket engine full of bolted-on subsystems when one elegant unit does the job? 💡
That’s what we did for segmentation.
✅ Meet the Encoder-only Mask Transformer (EoMT): tue-mps.github.io/eomt (CVPR 2025)
(1/6)

Went outside today and thought this would be perfect for my first #bluesky post