Meeting date: 20-06-2023 (10h30)
I contacted the authors of edgenet and asked for the pre-trained model, but they did not respond. As an alternative I found other pre-trained models where the output layer only has the mask of the drivable area.
I also added a new segmenting network (segformer head) trained on the ade20k dataset. This network has a relatively high inference speed, but was causing high latency. After some investigation I found that the delay was before the inference, but I can't explain why this happens. To avoid the problem I added a condition so that the inference only happens if the image delay is low.
For the evaluation of the models, I've been exploring the bdd100k dataset website. There are two distinct datasets, where one is used for object detection, drivable area and lane marking and the other for panoptic, semantic and instance segmentation. These two datasets overlap in some images, but with different names, so I was thinking of using a script that returns the equivalence between names of the two datasets - I've done it, but the result is an empty set .
Downloading the dataset files was very problematic, since the server uses a sort of tocken that expires before the download ends for larger files. As I understand it, this is only to ensure that the download is done by those who have accepted the terms and conditions. I tried deleting everything related to tockens and sessions from the link and it worked.
https://github.com/sithu31296/semantic-segmentation/blob/main/docs/MODELS.md
https://github.com/Qualcomm-AI-research/InverseForm
https://paperswithcode.com/sota/semantic-segmentation-on-cityscapes