This repository lists papers authored by Focoos AI.
| Title | Venue | Code |
|---|---|---|
| 📜 PEM: Prototype-based Efficient MaskFormer for Image Segmentation Niccolò Cavagnero, Gabriele Rosi, Claudia Cuttano, Francesca Pistilli, Marco Ciccone, Giuseppe Averta, Fabio Cermelli Prototype-based Efficient MaskFormer (PEM) is a transformer-based architecture for image segmentation that improves efficiency without sacrificing performance. It uses prototype-based cross-attention and a multi-scale feature pyramid network to reduce computation. PEM outperforms task-specific models while being more computationally efficient. |
CVPR 2024 | 🌐 Project Page  |
| 📜 The Revenge of BiSeNet: Efficient Multi-Task Image Segmentation Gabriele Rosi, Claudia Cuttano, Niccolò Cavagnero, Giuseppe Averta, Fabio Cermelli BiSeNetFormer is a multi-task image segmentation architecture designed for efficiency and accuracy, supporting semantic and panoptic segmentation. It combines two-stream architectures with a transformer-based segmentation head, achieving high inference speeds and competitive accuracy on datasets like Cityscapes and ADE20K. |
CVPR 2024 (Workshop) | - |
| 📜 What does CLIP know about peeling a banana? Claudia Cuttano, Gabriele Rosi, Gabriele Trivigno, Giuseppe Averta AffordanceCLIP leverages pre-trained Vision-Language models like CLIP to improve affordance segmentation for robots, bypassing the need for costly annotations or predefined actions. It achieves competitive zero-shot performance, works with any action prompt, and requires minimal additional training, enabling scalable, flexible models. |
CVPR 2024 (Workshop) | - |
| 📜 SAMWISE: Infusing Wisdom in SAM2 for Text-Driven Video Segmentation Claudia Cuttano, Gabriele Trivigno, Gabriele Rosi, Carlo Masone, Giuseppe Averta SAMWISE is a Referring Video Object Segmentation (RVOS) method that overcomes limitations of previous models by enabling streaming processing while retaining context. Built on the Segment-Anything 2 (SAM2) model, it integrates natural language understanding and temporal modeling, achieving state-of-the-art performance with minimal overhead. |
📝 Under submission |  |
Feel free to explore the papers and reach out for collaborations or inquiries!