Retraining TrafficCamNet primary detector with custom vehicle dataset

• Hardware (T4/V100/Xavier/Nano/etc) - Tesla T4
• Network Type (Detectnet_v2/Faster_rcnn/Yolo_v4/LPRnet/Mask_rcnn/Classification/etc) - Detectnet_v2
• TLT Version - 3.0

I am preparing a dataset for annotation in the process of retraining the trafficcamnet pretrained model for the class : car as the current model available in NGC doesn’t provide accuracy in night-time video streams for detecting vehicles.

I am preparing the dataset by annotating the RGB images. For the facenet training the dataset where 736*416 grayscale images.

Is there any particular image mode ( RGB; Grayscale; IR) that I should convert my whole dataset into and use to get higher accuracy on detections in night-time and low light areas.

Can you please help.

IR would be a solid option if your camera supports it. Feeding your net with a grayscale version of your dataset could be advantageous as it will force the net to not only learn based on the different channels (RGB) in your convolutions. I would first train the model with RGB save my weights and try again with grayscale. Depending on what your testing set gives you it should be pretty obvious whether grayscale helps. Don’t sweat with finding the perfect model too much (hint:you won’t) you can always do on the fly model updates.

@siozaris Thanks for the suggestion buddy. Will surely tryout your suggestion while testing the model.