karpathy / convnetjs

Deep Learning in Javascript. Train Convolutional Neural Networks (or ordinary ones) in your browser.
MIT License
10.89k stars 2.04k forks source link

Data input is important for perception algo of FSD? #125

Open xzg-wang-er opened 3 months ago

xzg-wang-er commented 3 months ago

Hello,

I'm very interested about the github course. and I'm wondoring about data input of the FSD autonomous driving algorithm perception solution, based on RGB image data as input or Camera raw as data input?

Here is some matetials on the AI day. On the 2021 AI day,it shows that Perception algo is based the camera raw data ,with the type "Bayer raw ,12bit ,no tonemapping". However ,On the 2022 AI day, it shows that perception algo is based the rgb(png/jpg) data ,with the type "rgb ,8bit ,isp processed".

My questions about the data input as follows: Q1:So the data input type description above is true ? what is the best data type for perception input. Q2:Shadow mode collects data such as Video clips after isp process and encode process,so the video image is the input for training? if not ,with the bayer raw as input,how to reduce the distribution gap between original raw data and rgb data。

Thank you very much. 20240807-102553 20240807-102548

371148606 commented 3 months ago

您好,我是薛文毅,已收到您的邮件,一会查收,谢谢!!