Transfer learning tool detectnet_v2.ipynp Visualize Inferences problem

Accelerated Computing Intelligent Video Analytics TAO Toolkit
1

Hello,

I have trained detectnet_v2 resnet10 model for 10 epoch, then when i try to visualize it gives %44 accuracy which is not bad for now ,however there is bboxes which does not cover any object, in my opinion its normal but their probabilities are too high and this seems problematic to me.

In the labels file which the visualizer tool creates there is 16 column as expected, 15 for training and 1 for accuracy of testing ( I could not find any official information about that but i read it from a forum ) and some of them have too high numbers, even if they located a object or not.

example:

ay 0.00 0 0.00 356.502 179.054 767.185 679.460 0.00 0.00 0.00 0.00 0.00 0.00 0.00 573.427

thanks

2

It is detectnet_v2_inference_kitti_tlt.txt may be required.

inferencer_config{

defining target class names for the experiment.

Note: This must be mentioned in order of the networks classes.

target_classes: “ay”

Inference dimensions.

image_width: 1200
image_height: 600

Must match what the model was trained for.

image_channels: 3
batch_size: 16
gpu_index: 0

model handler config

tlt_config{
model: “/workspace/tlt-experiments/detectnet_v2/experiment_dir_unpruned/model.step-13180.tlt”
}
}
bbox_handler_config{
kitti_dump: true
disable_overlay: false
overlay_linewidth: 2
classwise_bbox_handler_config{
key:“ay”
value: {
confidence_model: “aggregate_cov”
output_map: “ay”
confidence_threshold: 0.9
bbox_color{
R: 0
G: 255
B: 0
}
clustering_config{
coverage_threshold: 0.00
dbscan_eps: 0.3
dbscan_min_samples: 0.05
minimum_bounding_box_height: 4
}
}
}

classwise_bbox_handler_config{
key:“default”
value: {
confidence_model: “aggregate_cov”
confidence_threshold: 0.9
bbox_color{
R: 255
G: 0
B: 0
}
clustering_config{
coverage_threshold: 0.00
dbscan_eps: 0.3
dbscan_min_samples: 0.05
minimum_bounding_box_height: 4
}
}
}
}

3

You doubted the confidence value of the inerence result. There are two modes mentioned in tlt user guide. See Integrating TAO Models into DeepStream — TAO Toolkit 3.22.05 documentation