Trafficcamnet detect bicycle very low accuracy

I’m using Jetson Nano 4G with a downloaded pruned modle: trafficcamnet.

I did some simple testing by a logictech 720p camera, the targets are some web pages with object pictures, from my exp, can see the detections for people, car, roadsign are good, but bicycle looks really bad, any idea?

this is the config, i also tried camera-width, camera-height and fps to: 640, 480 and 1, same bad.
image

this is the primary infer config:

That’s possible. See below, the Two-Wheelers class has the least images. And also your test images may have different appearance against the two-wheelers training images.

From TrafficCamNet | NVIDIA NGC

TrafficCamNet v1.0 model was trained on a proprietary dataset with more than 3 million objects for car class. Most of the training dataset was collected and labeled in-house from several traffic cameras in a city in the US. The dataset also contains 40,000 images from a variety of dashcam to help with generalization and discrimination across classes. This content was chosen to improve accuracy of the object detection for images from a traffic cam at a traffic intersection.

Although persons, two-wheelers and road signs classes are included in the model, the accuracy of these classes will be much lower than car class. Some re-training will be required on these classes to improve accuracy.

BTW, please collect the training images which have similar data distribution as your testing images. That means it is better to trigger training with actual test images as much as possible.

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.