Accuracy not improving even after changing the input dim of DetectnetV2 Tao

Accelerated Computing Intelligent Video Analytics TAO Toolkit
,
1

Please provide the following information when requesting support.

• Hardware (T4/V100/Xavier/Nano/etc) : A40
• Network Type (Detectnet_v2/Faster_rcnn/Yolo_v4/LPRnet/Mask_rcnn/Classification/etc) : Detectnet_v2
• TLT Version (Please run “tlt info --verbose” and share “docker_tag” here) nvidia/tao/tao-toolkit: 5.5.0-pyt
• Training spec file(If have, please share here)
• How to reproduce the issue ? (This is for errors. Please share the command line and the detailed log here.)

I am doing model training using Detectnetv2 using Nvidia Tao on my custom dataset having one class with around 220 trainset 20 validationa nd 50 test set data points,

I trained the model on detectnetv2 using RN34(resnet34) backbone on Nvidia peoplenet input dimensions (960*544)and got the following result metrics

marker_1 AP 0.81818
mAP 0.818

Its same as when trained on the original detectnetV2 dimension(1248*384).

I cant understand no change in mAP because of input dimention change. I expect some change, it looks as if nothing changed !!

2

The DetectNet_v2 model, especially with a robust backbone like ResNet34, might be sufficiently robust to handle variations in input dimensions without significant performance degradation. This robustness could mean that the model effectively adapts to different resolutions, maintaining its detection capabilities.

More, your dataset is relatively small, with only 220 training images. In such cases, the model might not fully exploit the increased resolution or larger input size due to limited training data. Larger datasets often allow models to better utilize increased input dimensions.

The training process involves various hyperparameters (e.g., learning rate, batch size) that could influence how well the model adapts to different input sizes. If these parameters are not optimized for the new input dimensions, it might not lead to improved performance.

To explore potential improvements, consider the following steps:

  • Increase Dataset Size: Adding more diverse images could help the model better utilize higher resolution inputs.
  • Hyperparameter Tuning: Adjust learning rates, batch sizes, or other hyperparameters specifically for the new input dimensions.