Advances in Semantic Segmentation

The field of semantic segmentation is rapidly advancing with the development of new methods and techniques. One of the key trends is the use of foundation models, which are pre-trained on large datasets and can be fine-tuned for specific tasks. These models have been shown to achieve state-of-the-art performance on a variety of benchmarks, including the Pascal-5$^i$ and COCO-20$^i$ datasets. Another trend is the use of reinforcement learning, which has been used to enhance the pixel-level understanding and reasoning capabilities of large multimodal models. This approach has been shown to achieve remarkable performance on foreground segmentation tasks, such as camouflaged object detection and salient object detection. Additionally, there is a growing interest in using visual reference segmentation, which allows for the segmentation of objects without the need for manual annotations. Notable papers in this area include ProSAM, which introduces a probabilistic prompt encoder to improve the stability of visual reference segmentation, and ViRefSAM, which guides the Segment Anything Model using only a few annotated reference images. FastSeg is also a noteworthy paper, which proposes a novel and efficient training-free framework for open-vocabulary semantic segmentation. Overall, the field of semantic segmentation is rapidly advancing, with new methods and techniques being developed to improve performance and efficiency.

Sources

Few-Shot Segmentation of Historical Maps via Linear Probing of Vision Foundation Models

ProSAM: Enhancing the Robustness of SAM-based Visual Reference Segmentation with Probabilistic Prompts

Visual Content Detection in Educational Videos with Transfer Learning and Dataset Enrichment

Seg-R1: Segmentation Can Be Surprisingly Simple with Reinforcement Learning

FastSeg: Efficient Training-Free Open-Vocabulary Segmentation via Hierarchical Attention Refinement Method

Interactive Interface For Semantic Segmentation Dataset Synthesis

AI-Generated Lecture Slides for Improving Slide Element Detection and Retrieval

Foundation Models for Zero-Shot Segmentation of Scientific Images without AI-Ready Data

MammoTracker: Mask-Guided Lesion Tracking in Temporal Mammograms

NOCTIS: Novel Object Cyclic Threshold based Instance Segmentation

Mamba Guided Boundary Prior Matters: A New Perspective for Generalized Polyp Segmentation

Future Slot Prediction for Unsupervised Object Discovery in Surgical Video

ViRefSAM: Visual Reference-Guided Segment Anything Model for Remote Sensing Segmentation

Temporally-Aware Supervised Contrastive Learning for Polyp Counting in Colonoscopy

No time to train! Training-Free Reference-Based Instance Segmentation

Built with on top of