UniC-Lift: Rethinking 3D Scene Understanding with a Simpler and Unified Approach

Lifting two-dimensional segmentation masks into three-dimensional space is the standard approach for instance segmentation in neural radiance fields (NeRFs) and 3D Gaussian Splatting. The issue here is that these approaches require either pre- or post-processing techniques to estimate the final labels.

Rather than patching these inconsistencies with heavy, multi-stage post-processing, Ankit Dhiman, Srinath R, Jaswanth Reddy, Lokesh R Boregowda, and Venkatesh Babu Radhakrishnan propose UniC-Lift. This framework enforces consistency directly at the representation level. It assigns a learnable feature embedding to each 3D Gaussian and trains them via contrastive learning, pulling features of the same object together and pushing distinct objects apart. A triplet loss is also applied specifically to keep the object boundaries sharp.

What sets UniC-Lift apart is its embedding-to-label strategy. Rather than relying on expensive clustering (post-processing), labels are directly decoded from the learned embeddings through a simple thresholding mechanism. This strategy eliminates the requirement of a separate stage for clustering and simplifies the pipeline significantly.

By simplifying the pipeline into a single stage, UniC-Lift cuts training time while improving metrics on challenging benchmarks. In essence, UniC-Lift reframes 3D segmentation as a representation learning problem. By integrating everything into a single stage, it delivers efficient, scalable, and consistent scene understanding.


This work was presented at the AAAI Conference on Artificial Intelligence (AAAI 2026).

Project Page: https://unic-lift.github.io/

Paper: https://arxiv.org/abs/2512.24763

MENU