Sharing Test results with SceneSeg model -- foggy day during peak hour

[liuXinGangChina]

As a periodic demonstration of a long term data gathering and model validation task, we share the latest test result of model--ScensSeg that develop and maintain by privately-owned-vehicles working group. @m-zain-khawaja

This time we collect data from real high way, during morning peak hour on a foggy day.

1. link to 2Mp 120° FOV test result https://youtu.be/9IqOuMuZRg4 (1-minute)
2. link to 2Mp 30° FOV test result https://youtu.be/gz9hibMdb2g (1-minute)
3. link to 2Mp 30° FOV test result https://youtu.be/am8p5RSSNXk (15-minutes), including ramp-in, ramp-out and urban
4. link to 2Mp 120° FOV test result https://youtu.be/FMftopgY3Nc (15-minutes), including ramp-in, ramp-out and urban

Please feel free to leave your comment here no matter it is related to the data or the model.

We will continually sharing more test result.

[m-zain-khawaja]

Many thanks for collecting the data and sharing, most appreciated. Results seem quite stable with segmentation present for key on-road objects in both camera FoVs. I noticed that the truck on the right lane has some strange writing/markings, but it seems to not disturb the model segmentation since the lower part of the truck is always detected. When combined with RADAR, this should give a strong detection of objects.

[liuXinGangChina]

Hello Khawaja san,

1. the marking and paiting on the back is following chinese laws, because the plate may be covered by mud during driving
2. this is just short clip, we will release more when the post-process finished

Have a nice day!

心刚

[m-zain-khawaja]

Thanks for your message and clarifying - looking forward to seeing more clips once they are finished processing 👍

[liuXinGangChina]

khawaja san, @m-zain-khawaja

Already add the link for longer version (15 minutes), please check
feel free to leave your comment when you have question about the content

Best regards！

心刚

[m-zain-khawaja]

Thank you very much for sharing the longer versions - great to see this testing in China, with a wide variety of vehicle types and trucks/buses with interesting markings being nicely detected in both FoVs.

I still think that for the Wide FOV camera, there is potentially some interference from ego-car bonnet pixels on detections for other vehicles in the left/right lane when they are very close to the ego-car. Is your pre-processing pipeline as follows:
-> Capture Wide FoV Image
-> Remove bonnet pixels by cropping out bottom of image
-> Crop evenly from left/right of image to achieve a 2:1 width to height aspect ratio
-> Run network

[liuXinGangChina]

Hello， khawaja san, @m-zain-khawaja

It is reasonable to crop the picture to "2:1 width to height aspect ratio"
will inference again after preprocess the wide input

Have a nice day

心刚