ParallelLLC/Segmentation

1---2language:3- en4tags:5- computer-vision6- segmentation7- few-shot-learning8- zero-shot-learning9- sam210- clip11- pytorch12license: apache-2.013datasets:14- custom15metrics:16- iou17- dice18- precision19- recall20library_name: pytorch21pipeline_tag: image-segmentation22---23 24# Model Card for SAM 2 Few-Shot/Zero-Shot Segmentation25 26## Model Description27 28This repository contains two main models for domain-adaptive segmentation:29 30### SAM2FewShot31- **Architecture**: SAM 2 + CLIP with memory bank32- **Purpose**: Few-shot learning for segmentation33- **Input**: Images + support examples34- **Output**: Segmentation masks35 36### SAM2ZeroShot  37- **Architecture**: SAM 2 + CLIP with advanced prompting38- **Purpose**: Zero-shot learning for segmentation39- **Input**: Images + text prompts40- **Output**: Segmentation masks41 42## Intended Uses & Limitations43 44### Primary Use Cases45- Domain adaptation for segmentation tasks46- Rapid deployment in new environments47- Minimal supervision scenarios48- Research in few-shot/zero-shot learning49 50### Limitations51- Performance depends on prompt quality52- Domain-specific adaptations required53- Computational cost of attention mechanisms54- Limited cross-domain generalization55 56## Training and Evaluation Data57 58### Domains59- **Satellite Imagery**: Buildings, roads, vegetation, water60- **Fashion**: Shirts, pants, dresses, shoes61- **Robotics**: Robots, tools, safety equipment62 63### Evaluation Metrics64- IoU (Intersection over Union)65- Dice coefficient66- Precision and Recall67- Boundary accuracy68- Hausdorff distance69 70## Training Results71 72### Few-Shot Performance (5 shots)73| Domain | Mean IoU | Mean Dice |74|--------|----------|-----------|75| Satellite | 65% | 71% |76| Fashion | 62% | 68% |77| Robotics | 59% | 65% |78 79### Zero-Shot Performance (Best Strategy)80| Domain | Mean IoU | Mean Dice |81|--------|----------|-----------|82| Satellite | 42% | 48% |83| Fashion | 38% | 45% |84| Robotics | 35% | 42% |85 86## Environmental Impact87 88- **Hardware Type**: GPU (NVIDIA V100 recommended)89- **Hours used**: Variable based on experiments90- **Cloud Provider**: Any cloud with GPU support91- **Compute Region**: Any92- **Carbon Emitted**: Depends on usage93 94## Citation95 96```bibtex97@misc{sam2_fewshot_zeroshot_2024,98  title={SAM 2 Few-Shot/Zero-Shot Segmentation: Domain Adaptation with Minimal Supervision},99  author={Your Name},100  year={2024},101  url={https://huggingface.co/esalguero/Segmentation}102}103```104 105## Model Card Authors106 107This model card was written by the research team.108 109## Model Card Contact110 111For questions about this model card, please contact the repository maintainers.