Custom model nvinfer facedetection with wrong output problem

Deepstream 6.2
ubuntu 22.04

Problem: I'm encountering an issue with the model implemented using NVinfer for face detection, as the output appears to be incorrect. I'm seeking guidance on addressing this issue and improving the accuracy of the model's face detection capabilities. I would greatly appreciate your assistance in providing insights or suggestions to rectify the incorrect output and enhance the model's performance.

Desctiption: Using Onnx Model(https://github.com/onnx/models/tree/69c5d3751dda5349fd3fc53f525395d180420c07/vision/body_analysis/ultraface) - version-RFB-640

Onnx-output Information:
## nput

### Original model

Image, name - `input`, shape - `1, 3, 240, 320`, format `B, C, H, W`, where:

* `B` - batch size
* `C` - number of channels
* `H` - image height
* `W` - image width

Expected color order is `RGB`.

Mean values - [127.0, 127.0, 127.0]. Scale values - [128.0, 128.0, 128.0].

### Converted model

Image, name - `input`, shape - `1, 3, 240, 320`, format `B, C, H, W`, where:

* `B` - batch size
* `C` - number of channels
* `H` - image height
* `W` - image width

Expected color order is `BGR`.

## Output

### Original model

1. Bounding boxes, name: `boxes`, shape - `1, 4420, 4`. Presented in format `B, A, 4`, where:
  * `B` - batch size
  * `A` - number of detected anchorsFor each detection, the description has the format: [`x_min`, `y_min`, `x_max`, `y_max`], where:
  * (`x_min`, `y_min`) - coordinates of the top left bounding box corner (coordinates are in normalized format, in range [0, 1])
  * (`x_max`, `y_max`) - coordinates of the bottom right bounding box corner (coordinates are in normalized format, in range [0, 1])
2. Scores, name: `scores`, shape - `1, 4420, 2`. Contains scores for 2 classes - the first is background, the second is face.

### Converted model

1. Bounding boxes, name: `boxes`, shape - `1, 4420, 4`. Presented in format `B, A, 4`, where:
  * `B` - batch size
  * `A` - number of detected anchorsFor each detection, the description has the format: [`x_min`, `y_min`, `x_max`, `y_max`], where:
  * (`x_min`, `y_min`) - coordinates of the top left bounding box corner (coordinates are in normalized format, in range [0, 1])
  * (`x_max`, `y_max`) - coordinates of the bottom right bounding box corner (coordinates are in normalized format, in range [0, 1])
2. Scores, name: `scores`, shape - `1, 4420, 2`. Contains scores for 2 classes - the first is background, the second is face.

We are moving this post to the Deepstream forum to get better help.

There is no update from you for a period, assuming this is not an issue anymore. Hence we are closing this topic. If need further support, please open a new one. Thanks

@sweum Can you tell us what problem on earth you met?

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.