Percentile Clipping based Low Bit-Precision Quantization for Depth Estimation Network
Seungeon Hwang, Jongsun Park · 2022 19th International SoC Design Conference (ISOCC) · 2022
Low bit-precision quantization is an efficient method of reducing the computation complexity of neural networks. However, low bit-precision quantization of complex tasks such as depth estimation is a challenging issue due to the multiple skip connections of U-Net like networks. To maintain accuracy while lowering bit width, activation statistics should be considered. In this paper, we propose a percentile clipping technique for low bit-precision quantization for depth estimation network. By selecting the clipping point as the percentile value considering the distribution of the skip connection parts, our proposed technique can efficiently reduce the relative squared quantization error by 34.78% when using uniform 4-bit weights and activations.