search

irtrailenium / train-ego-path-detection

MIT License
16 stars 2 forks source link

Hello authors, can you provide a weight file for training with Res-net34. #1

Open MrG27-web opened 2 weeks ago

MrG27-web commented 2 weeks ago

Hello authors, can you provide a weight file for training with Res-net34。

GreatTommy commented 1 week ago

Hello MrG27-web, Unfortunately, we don't provide trained weights for the ResNet-34 backbone as part of this work. However, you can easily reproduce the training process by following the instructions in the README file. If you have any questions or need further assistance, feel free to ask!

MrG27-web commented 1 week ago

Hello MrG27-web, Unfortunately, we don't provide trained weights for the ResNet-34 backbone as part of this work. However, you can easily reproduce the training process by following the instructions in the README file. If you have any questions or need further assistance, feel free to ask!

Hello Thomas,

Thank you for your reply. When performing video inference, I have found that uncertainty increases when a train passes through a switch. To address this issue, I have considered a strategy: using the track regression results from the previous frames to crop the next frame's image, thereby hiding the surrounding tracks, and using the cropped image for subsequent inference. In this way, the model focuses only on the area that the train is likely to pass through during inference. However, I have encountered a problem: while it has some effect at the beginning, over time, the errors from this cropping method accumulate, leading to the subsequent cropping hiding the correct areas and an inability to correctly recognize the tracks.

Do you have any suggestions for this approach?