Rare Sense Inc.

Comparative Evaluation of SAM1 and SAM2 Over the past month, I’ve been deeply focused on a critical challenge in the world of fashion AI: creating precise masking techniques for clothing. Accurate masking is essential for the next steps in our pipeline, such as generating and warping new outfits on models. However, the lack of a reliable model for this task has been a consistent bottleneck for many engineers in the field. With the release of Segment Anything Model 2 (SAM 2) by Meta on July 29th, I was excited to explore its potential, especially with the claims of improved segmentation. However, after rigorous testing, I found that while SAM 2 excels in speed, particularly for video segmentation tasks (thanks to its use of bfloat16 precision), it struggles with image segmentation. I revisited Segment Anything Model 1 (SAM 1) and found that it outperformed SAM 2 in both dot prompts and box prompts. This allowed me to achieve significantly better results in image segmentation. You can view the comparative results with dot prompts https://lnkd.in/dQBAEA7x and box prompts (Grounding Dino use for cloth detection) https://lnkd.in/d2aGYzmg Special thanks to my senior, Muhammad Bilal,(Sir apka vision hai ????♂) for helping me navigate through these experiments and find the best model for the task. Together, we are also fine-tuning both SAM1 and SAM 2 to optimize them further for fashion applications. Stay tuned, as I will soon be sharing a detailed and easily understandable training code for both dot and bounding box prompts! ?? The journey towards better fashion AI continues—In Sha Allah! ?? . . . #fasionai #llm #sam1 #sam2 #sam #segmentanything #segmentation #ai #generativeai #cv #meta

Comparative Evaluation of SAM1 and SAM2 Over the past month, I’ve been deeply focused on a critical challenge in the world of fashion AI: creating precise masking techniques for clothing. Accurate masking is essential for the next steps in our pipeline, such as generating and warping new outfits on models. However, the lack of a reliable model for this task has been a consistent bottleneck for many engineers in the field. With the release of Segment Anything Model 2 (SAM 2) by Meta on July 29th, I was excited to explore its potential, especially with the claims of improved segmentation. However, after rigorous testing, I found that while SAM 2 excels in speed, particularly for video segmentation tasks (thanks to its use of bfloat16 precision), it struggles with image segmentation. I revisited Segment Anything Model 1 (SAM 1) and found that it outperformed SAM 2 in both dot prompts and box prompts. This allowed me to achieve significantly better results in image segmentation. You can view the comparative results with dot prompts https://lnkd.in/dQBAEA7x and box prompts (Grounding Dino use for cloth detection) https://lnkd.in/d2aGYzmg Special thanks to my senior, Muhammad Bilal,(Sir apka vision hai ????♂) for helping me navigate through these experiments and find the best model for the task. Together, we are also fine-tuning both SAM1 and SAM 2 to optimize them further for fashion applications. Stay tuned, as I will soon be sharing a detailed and easily understandable training code for both dot and bounding box prompts! ?? The journey towards better fashion AI continues—In Sha Allah! ?? . . . #fasionai #llm #sam1 #sam2 #sam #segmentanything #segmentation #ai #generativeai #cv #meta