Mean average precision of 0.00 in training Trafficcamnet model using Tao Toolkit

Accelerated Computing Intelligent Video Analytics TAO Toolkit
22

I did it already and also set the
minimum_bounding_box_height to lower, for example, 4.
minimum_height to lower, for example, 4.
minimum_width to lower, for example 4
but mistakely may be send older spec file

this is latest one
spec.txt (6.4 KB)

23

OK, please monitor the evaluation result.
More, please use the latest spec file to train to the end. It is normal to get lower mAP for the 1st epoch.

24

Evaluation complete still same log…
validation.log (47.0 KB)

25

You are evaluating against the 1st model: -m /home/trainval/latest_training/model/model.step-0.tlt

As mentioned above, please run training to the end. You can monitor the mAP result during training or the end of training. It is normal to get lower mAP for the 1st epoch.

26

but mAP is coming 0.00

Mean average_precision (in %): 0.0000

class name       average precision (in %)
-------------  --------------------------
four_wheeler                            0
heavy                                   0
three_wheeler                           0
two_wheeler                             0

Median Inference Time: 0.008560
2025-01-13 07:55:33,330 [INFO] __main__: Evaluation complete.
Time taken to run __main__:main: 0:41:49.256878.

is this normal atleast average precision also have to come in few numbers

Another thing i have to ask that the model i am using for training is pretained model have classes car, road-sign, person
now i trained on four_wheeler, heavy, three_wheeler, two_wheeler
and when i use this model still detect classes car, road-sign, person?

27

It is normal to get mAP0 for the 1st epoch.

Please make sure the groundtruth are correct. The network can detect your new classes.

More, resnet18 may not be enough to get better result. You can run more experiments with deeper backbones, finetuning parameters,etc.

Also, you can use other networks(yolo_v4 tiny, DINO) to train against this task.