Edge-Intelligence

1. For edge-camera, with high-speed inference. Mainly solves the problem of tight computility (computing power) on the edge side in industrial scenarios.

   1. SenseTime research: dynamic computility model, and model selection. We aim to apply less consumption in the inference processing (Based on every single input sample, to choose a model with fewer parameters, or stopping early in a large model)
   2. Technicals: MOE (Mixture of Experts), Early stop, Ensemble learning

2. For edge-VR devices, with high storage consumption of volumetric video.

   1. Compressed Volumetric Video Size (More on computer vision)
   2. Accelerated transmission over the network
   3. Edge-side cache

3. For edge-3D cameras, a new paradigm in smart city.

   1. Multi-camera tracking
   2. Cloud-Edge Collaborative Computing (Tasks are accomplished on the edge side as much as possible, and complex tasks are accomplished efficiently on the cloud side.)

Future area:

1. For edge devices for LLM, LLM adaptation.