Negative instances/images for Detectnet_v2 in TLT

I have been training detectnet_v2 models for sometime now in using TLT. I have high recall for my classes. But I get a lot of false positives when similar objects are in the scene. I wish to add negative instances so that my model learns how to distinguish between the actual object and its lookalike. How do I add these instances? Do I add images with only these instances and no other classes i.e. do I provide images with empty kitti labels? or do I need to have atleast one object from one of my target classes with the negative instances? Does detectnet_v2 in TLT accept empty kitti labels? Any help will be appreciated.

TLT does not accept empty kitti labels.
For negative samples, refer to How to give negative data-set while training model on TLT - #6 by Morganh

So as I have understood Detectnet does not do auto negative sampling like FRCNN and SSD. Means anything other than the ground truth boxes are not taken as background? So does that also mean that adding negative samples will not increase my precision? I will keep getting false positives with similar objects?

Just to clarify I am not talking about labelling similar looking objects as a separate class. I simply want to have those objects in my images along with my target classes. But I will only label the classes I require.

Detectnet_v2 does not do auto negative sampling. If you want to detect similar objects, please generate their labels in the training dataset and then let detectnet_v2 train them as a new class too.
BTW, how many classes did you train?

I have 4 classes. Face. Mask. Gun. Helmet. My models keep detecting anything black that is near or in the hands as gun (like a mobile phone) or any L shaped object as gun (like the legs of an office chair or even sometimes a person’s legs when wearing black jeans) and a full head of hair as helmet. Face and mask are not creating issues as they are mutually exclusive. The recall of all the classes are high. But lots of false positives for gun and helmet. So any ideas on preventing false positives for Detectnet_v2?

My suggestions:

  1. Please double check the Gun/Helmet labels of all the images. Make sure they are expected.
  2. 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.

  1. Refer to
    Too many false positives. - #11 by Morganh
    PeopleNet precision low for person class - #23 by Morganh
    Which detection model will give more accuracy for arial view image detection! - #2 by Morganh