Results from SSD models for ITS

We have trained a model for detecting and classifying 170 Make models of cars and 10 types of vehicles
thought it will be interesting to share.
http://139.59.121.84/ssd_vehicle_classification/

Articulated truck
Bicycle
Bus
Car
Motorcycle
Motorized vehicle (i.e. Vehicles that are too small to be labeled into a specific category, Trike, Rickshaw)
Non-motorized vehicle
Pedestrian
Pickup truck
Single unit truck
Work van


Thanks for the sharing! Looks cool.

Could you also share your hardware/framework/performance information?
Thanks.

Hi AastaLLL,

Results are from Jetson TX1 based camera. Used for parking encroachment and violation detection.

batch_size,mean processing time,mean time per img
1,214.4,214.4
2,411.2,205.6
4,815.8,203.95
8,1604.2,200.525
16,3256.0,203.5

Camera details
https://devtalk.nvidia.com/default/topic/1027276/announcements/rugged-jetson-based-360-camera/


Few more images, from live cam

Very impressive Ravi, are your models available with the camera and also separately?

What dataset did you use to train vehicle model?

Hi Dusty,

For now models are available with camera.We are collecting data from local municipalities.
Will arrange an live feed, once we get permission.

what approximate framerate are you processing the images at on TX1 with SSD? Is ‘mean time per image’ milliseconds?

Yes. This particular project is for parking violation detection. So the model can optimised for speed

The night-time performance is quite impressive. Thanks for sharing.

what approximate framerate are you processing the images at on TX1 with SSD? Is ‘mean time per image’ milliseconds?

hi, there. I wanna know which gie your work use, nvinfer or gst-plugin just like yolo-plugin nvidia published on github?
And can it work with multi sources of video as multi-channels? I’d like to display videos as a 2d-array as the same time, like samples deepstream-app does.
Thanks a lot.

Live demo is here.

//

@mh_ding we have an custom inference engine, yes it works with multiple inputs

Very impressive Ravi, are your models available with the camera and also separately?

What dataset did you use to train vehicle model?

Please the video below, runs live on TX2