Open World Object segmentor

Description

I am developing a lightweight, real-time open-world segmentation system designed to break images into fine primitive regions before forming final object candidates. Instead of attempting to predict complete object masks in a single pass, the system follows an “over-segment first, group later” approach. This preserves more structure early in the pipeline and gives downstream stages a stronger foundation for object discovery, recognition, and scene understanding.

The core idea is that small regions can be merged later, but object parts that are merged too early are difficult to separate again. By maintaining fine scene detail at the beginning of the pipeline, the system is better suited for open-world environments where objects may not belong to known training categories. This makes the approach useful for robotics, autonomous navigation, and real-time perception systems that need to reason about unfamiliar objects.

The current pipeline focuses on class-agnostic scene decomposition, primitive region generation, boundary preservation, and hierarchical merging. The long-term goal is to build a fast, practical alternative to large segmentation systems like SAM for applications where real-time performance and deployment efficiency matter.

Key Result

The current system reaches 10 FPS on an RTX 4070 Laptop GPU with 8 GB of VRAM, meeting my initial real-time performance target.