Segment Anything Model (SAM) Meets Glass: Mirror and Transparent Objects Cannot Be Easily Detected

A key factor that drives the development of generative AI is foundation model Bommasani et al. [2021] that at inference can generalize to tasks and data distributions different from training. With the success of ChatGPT Zhang et al. [2023b], GPT-3 [Brown et al., 2020] has been widely recognized as one of the most widely recognized foundation models for NLP. Very recently, Meta AI research team has recent released a segment anything project Kirillov et al. [2023] that introduces a promotable segmentation task for training a vision foundation model. The resulting segment anything model (SAM) has been recognized as the GPT-3 moment for vision. The model was trained on over 1 billion masks on 11 million licensed and privacy-respecting images. It represents a significant step towards achieving cognitive recognition for all objects in the world, aiming to handle interactive segmentation tasks while addressing real-world constraints.