Case Study: Texas Instruments achieves real-time semantic segmentation on TDA4VM with Embedl

Texas Instruments achieves 3x faster semantic segmentation with Embedl

Texas Instruments faced the challenge of deploying a real-time semantic segmentation model for autonomous driving applications on their resource-constrained TDA4VM edge processor. The initial Lightweight Atrous Spatial Pyramid Pooling (LRASPP) model's architecture was not natively compatible with TI's toolchain, and achieving both low latency and high accuracy was difficult. They partnered with Embedl to address this.

Using its Model Optimization SDK, Embedl adapted the model for compatibility and then applied its latency-aware pruning techniques. The optimized solution delivered a 3x speedup, reducing latency from 46.76 ms to 15.5 ms on the TDA4VM hardware while maintaining near-original accuracy, with only a minimal drop in mIoU from 53.3% to 52.4%. This enabled Texas Instruments to run accurate, real-time semantic segmentation on their power-efficient platform.

Texas Instruments