Dequantisation issue in fast depth model

General
, , , ,
1

I am trying to run some inference using the NYUv2 dataset for depth estimation using the fast-depth.hef pre compiled model which i downloaded from Model Explorer Vision.

I tried to run the inference using just one sample image.

The output tensor i am getting is not matching with the ground truth tensor.

OUTPUT TENSOR
Output Tensor shape: (224, 224)
Output Tensor min/max: 1.659901738166809/9.000823974609375
Output Tensor dtype: float32
Output Tensor

[[3.45439 3.45439 3.9497056 … 3.9890745 3.7684255 3.7684255]
[3.45439 3.45439 3.9497056 … 3.9890745 3.7684255 3.7684255]
[3.8160343 3.8160343 4.618061 … 4.193243 3.9579456 3.9579456]
…
[1.9748528 1.9748528 2.0929594 … 1.7624439 1.6992706 1.6992706]
[1.8411816 1.8411816 2.0059817 … 1.6681417 1.6599017 1.6599017]
[1.8411816 1.8411816 2.0059817 … 1.6681417 1.6599017 1.6599017]]

GROUNDTRUTH TENSOR

GT Tensor shape: (224, 224)
GT Tensor min/max: 2.1436519622802734/9.318137168884277
GT Tensor dtype: float32
GT Tensor

[[4.938948 4.9406 4.948493 … 6.511611 6.5083556 6.506549 ]
[4.9373155 4.9389176 4.946815 … 6.51131 6.5079646 6.506164 ]
[4.9329944 4.9343886 4.9418097 … 6.5105906 6.5068154 6.5051317]
…
[2.6368117 2.636188 2.6340883 … 2.4576383 2.7947624 2.797908 ]
[2.6317499 2.630723 2.6248417 … 2.7684326 2.7861063 2.7919548]
[2.6283824 2.6271293 2.6207948 … 2.7672853 2.7810218 2.7879908]]

This is the preprocess function i am using.

def preprocess_h5_for_hailo(h5_file):

with h5py.File(h5_file, 'r') as f:

    rgb = np.array(f\['rgb'\])

    if rgb.shape\[0\] == 3:   # channel-first → convert to HWC

            rgb = np.transpose(rgb, (1, 2, 0))



    \# Resize to model input (224x224)

    img = cv2.resize(rgb, (224, 224))

    return np.expand_dims(img, axis=0)

Also i have set the quantised parameter for the input vstream as True and for the output vstream it’s false.

input_vstreams_params = hpf.InputVStreamParams.make_from_network_group(network_group, quantized=True)

    output_vstreams_params = hpf.OutputVStreamParams.make_from_network_group(network_group,quantized=False, format_type=hpf.FormatType.FLOAT32)

Am i doing things correctly?
If yes, why is my output tensor not matching with the ground truth tensor.

Here is the link to the entire code that i am using. Code