Imagine having to point at every single car in a thousand photos just so an AI can recognize them. Sounds tedious, right? This is where Dolby’s Image Segmentation becomes crucial.

Normally, we enable a Segment Anything Model (SAM) that can isolate almost any object in an image, which needs a human prompt every single time. But doing so, the AI becomes inefficient when you need to prompt endlessly to identify each object repeatedly for N objects per image.

T-SAM Is Teaching AI to See on Its Own

Researchers at Dolby built on a technique called transductive learning to address this issue. Instead of retraining the model or prompting it repeatedly, T-SAM studies a handful of labeled images and automatically generates prompts for new ones — without repeated human input

T-SAM explores two approaches:

The first averages patterns from labeled examples into a single general instruction.

The second compares each new image against known examples and builds a tailored prompt on the fly, adapting to variation in objects and scenes.

When tested across multiple segmentation benchmarks, T-SAM consistently improved as we provided more labelled examples. Its accuracy improved 3x more between one-shot and five-shot settings, nearly matching the performance of SAM2 when tested across common objects in context. The model also delivered standout results on several object categories, achieving up to 19× higher accuracy than baseline approaches in some one-shot scenarios. Most importantly, the experiments showed that T-SAM continues to learn from user feedback, becoming increasingly effective after just a handful of labelled examples.

Looking ahead, T-SAM points toward AI vision tools that make large-scale image analysis faster, cheaper, and far more practical across industries from media to healthcare. It reduces labeling effort significantly, especially in situations where there are many similar objects or large datasets.

Smarter Image Segmentation with T-SAM: Less Prompting, More Precision.

{{ pageEducationalSlot.headingContent[index] }}