The problem of reconstructing the dataset

[carryman97]

Hello, I tried to create a dataset, but I encountered a problem. I don't know how to reconstruct the two dataset files "right_dataset" and "left_dataset". When I did not use these two dataset files to train the model agent, I found that the agent turned left or right at the intersection very poorly and could not correctly drive into the prescribed lane. I hope you can help me solve the problem. What should I do to rebuild the two dataset files "right_dataset" and "left_dataset"? Thank you!

[Kait0]

See https://github.com/autonomousvision/transfuser/issues/188. However TransFuser should be able to make turns consistently even without these routes so I suspect your problem lies elsewhere.

[carryman97]

I reconstructed the dataset based on your open source data collection code, and when I reconstructed the dataset, each route was randomly assigned a type of weather (sunny, cloudy, rainy, etc.), light level, sun azimuth, etc., instead of Randomly assign these parameters to each frame (because I want to change the model input from a single frame input to a time series frame input. If the scenes of adjacent frames are different, then there will be no correlation between the preceding and following frames. For this I Need to rebuild the dataset). When I collected the data and used it to train the same model, I got bad results. I want to ask you, did you clean the data after collecting the data? For example, delete data where collisions occurred, or delete data where the route was not completed. Or do you have any other good suggestions? The problem of data sets has troubled me for a long time.Thank you！

crazyman @.***

 

------------------ 原始邮件 ------------------ 发件人: "autonomousvision/transfuser" @.>; 发送时间: 2023年12月4日(星期一) 晚上10:19 @.>; @.**@.>; 主题: Re: [autonomousvision/transfuser] The problem of reconstructing the dataset (Issue #198)

See #188. However TransFuser should be able to make turns consistently even without these routes so I suspect your problem lies elsewhere.

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

[Kait0]

Bad result is too vague to give you any concrete advise. We didn't do any particular clearning in this repository. I would first make sure your training pipeline is working properly. Train with the original dataset and see if you can reproduce the result. If not than there is some problem with the training setup.

Also I wanted to mention, in our latest repository for TransFuser++ we use a dataset that changes the weather etc. in every route (instead of every frame), like you need.