Retraining DashcamNet with TLT

DashcamNet has four objects detection. Now I am going to use it for vehicle detection.
The model is good for side view detection, but not good for topdown view.
I am going to retrain with images for topdown view.
I am going to retrain for topdown view vehicles only.
Is there any concern for model accuracy? Since model is trained with four classes and I am going to retrain for only one class.
Is this ok to retrain?
What are the things I need to caution in retraining?

It is ok to retrain with only one class.
For topdown view objects, they may be small.
Refer to Frequently Asked Questions — Transfer Learning Toolkit 3.0 documentation

In DetectNet_V2, are there any parameters that can help improve AP (average precision) on training small objects?

Following parameters can help you improve AP on smaller objects:

  • Increase num_layers of resnet
  • class_weight for small objects
  • Increase the coverage_radius_x and coverage_radius_y parameters of the bbox_rasterizer_config section for the small objects class
  • Decrease minimum_detection_ground_truth_overlap
  • Lower minimum_height to cover more small objects for evaluation

Yes the size is small, any issue for that?

For small objects, you can refer to above guide mentioned in Frequently Asked Questions — Transfer Learning Toolkit 3.0 documentation

Previously I used TLT to train detectnet_v2 with Resnet18 for one object detection.
But now I am going to retrain DashcamNet.
How about image size to retrain DashcamNet.
What is the image size I need to use for retraining?

Please set
output_image_width: 960
output_image_height: 544

and setting the enable_auto_resize parameter to true

Reference: DetectNet_v2 — Transfer Learning Toolkit 3.0 documentation

Thanks a lot

This topic was automatically closed 60 days after the last reply. New replies are no longer allowed.