what's new in computer vision

Yupeng Zheng, Kai Zou, Bin Liu +1

this paper proposes visco, a method that uses large language models as intrinsic encoders to compress visual tokens. this is important for practitioners working with vision-language models, as it can significantly reduce the computational cost and memory usage associated with processing large numbers of visual tokens.

Eleftherios Tsonis, Xi Wang, Vicky Kalogeiton

this research investigates learning olfactory representations guided by language, acknowledging that images alone don't capture all sensory information. it offers a new direction for multimodal ai, allowing practitioners to build systems that understand and reason about non-visual sensory data.

Yiteng Yuan, Junyan Wang, Zheyuan Liu +4

this paper proposes a flow matching method for clustering data where information is incomplete across different views. it helps practitioners analyze complex datasets with missing modalities, improving clustering performance in real-world scenarios where data collection is imperfect.

Zheng Guo, Jiaqi Cui, Haocheng Xiong +5

koal proposes a knowledge-driven approach for prostate cancer grading using ordinal-aware learning. this method provides a more accurate and clinically relevant assessment from medical images, which is valuable for practitioners in developing diagnostic ai tools.

Guang Yang, Brian Siyuan Zheng, Victoria Ebert +1

this paper introduces legato 2, a pipeline designed to extract symbolic musical notation from sheet music. it combines visual recognition with musical understanding, which is useful for digitizing scores, aiding music analysis, or creating interactive musical applications.

Anthony Hu, Václav Volhejn, Adrien Ramanana Rahary +7

this work introduces multiplayer interactive world models using representation autoencoders to simulate complex interactive environments. for practitioners, this enables the creation of more sophisticated and realistic virtual worlds for training agents or developing interactive applications.

Riccardo Renzulli, Gabriele Spadaro, Shruthi Gowda +2

torino introduces a method to reduce the number of tokens in vision-language models by identifying and merging redundant visual concepts. this can make large vlm inference more efficient and interpretable, which is practical for deploying these models in resource-constrained environments.

Francesca Morandi, Omayma Moussadek, Federico Venturini +5

this paper proposes using task arithmetic to achieve robust zero-shot generalization in open-vocabulary action recognition. for practitioners, this means developing models that can recognize new actions without requiring explicit training examples, making action recognition systems more adaptable and scalable to previously unseen scenarios.

Jan Glaser, Ivo Bukovsky, Noriyasu Homma +1

this paper extends the concept of learning entropy to image analysis, creating 'entropy signatures' for representation and classification. practitioners can use this novel method to extract more robust and distinctive features from images, potentially enhancing performance across various image classification tasks.

Aryan Das, Koushik Biswas, Moloud Abdar +1

this paper introduces unity, an adapter for efficient and scalable composite conditioning in diffusion-based image generation. for researchers and developers working with diffusion models, unity provides a method to improve the flexibility and control over generated outputs by adaptively integrating various conditioning signals.

Yohann Perron, Guillaume Astruc, Nicolas Gonthier +2

this paper presents universat, a transformer model capable of processing satellite imagery regardless of its resolution or sensor modality. earth observation practitioners can leverage this to develop more flexible and generalizable models, reducing the need for distinct architectures for varying data types and resolutions.

Reeshad Khan, John Gauch

this work investigates optimizing camera sensor design, moving past traditional bayer filters, specifically for autonomous driving segmentation tasks. for practitioners, this means the potential to achieve more robust perception in self-driving vehicles by jointly designing sensors and algorithms, improving performance in challenging conditions.