This is the project page for the ICCV 2023 paper:

4D Panoptic Segmentation as Invariant and Equivariant Field Prediction

Our model ranks 1st on the SemanticKITTI 4D Panoptic Segmentation public leaderboard!

Two models are presented in our paper: EQ-4D-STOP and EQ-4D-PLS.

Code open-sourced!

EQ-4D-StOP: this link.

EQ-4D-PLS: coming soon.

TL;DR

Main idea

We view the instance clustering strategies through the lens of invariant and equivariant field regression, yielding the two models.

Qualitative result

Take EQ-4D-StOP as an example, the regressed offsets from the equivariant model more accurately point to the instance centers.

Quantitative result

Our equivariant models achieve higher accuracy (LSTQ) with lower memory consumption and running time compared with the non-equivariant baseline.