I created a NN that subtracts 2 depth frames: class DiffImgs(nn.Module): def forward(self, img1, img2): if img1.shape != (1, 1, 960, 1632) or img2.shape != (1, 1, 960, 1632): raise ValueError('Input images must have shape (1, 1, 960, 1632)') return torch.sub(img1, img2) I get the output as a vector<float> by using: auto depth_diff_message = queues["depth_diff"]->get<dai::NNData>(); std::vector<float> depth_diff_data = depth_diff_message->getFirstLayerFp16() The input of StereoDepth with subpixel to the model is UINT16 The model does the diff and outputs at Fp16 Because I subtracted it means the results can be negative or positive getFirstLayerFp16 outputs a 1D vector of floats In the depthai C++ repo it looks like there needs to be some conversion from NNData to get into a CV frame: https://github.com/luxonis/depthai-core/blob/main/examples/utility/utility.cpp cv::Mat fromPlanarFp16(const std::vector<float>& data, int w, int h, float mean, float scale){ cv::Mat frame = cv::Mat(h, w, CV_8UC3); for(int i = 0; i < w*h; i++) { auto b = data.data()[i + w*h * 0] * scale + mean; frame.data[i*3+0] = (uint8_t)b; } for(int i = 0; i < w*h; i++) { auto g = data.data()[i + w*h * 1] * scale + mean; frame.data[i*3+1] = (uint8_t)g; } for(int i = 0; i < w*h; i++) { auto r = data.data()[i + w*h * 2] * scale + mean; frame.data[i*3+2] = (uint8_t)r; } return frame; } What is the proper way to convert a diff'd StereoDepth frame from vector<float> into cv::Mat?

Hi AdamPolak , I don't think that's possible - afaik NN can only output FP16. It's the same, you are expecting depth (INT16) anyways, not disparity (where it changes from INT8 to INT16 when subpixel is enabled).

erik I subtract the depth frame from 0's in the model: So it is just outputting the same depth frame values. I get the the NNData from the queue and check that the size is the same as the expected 1632x960 and it is I create a cv::Mat and then iterate through the floats, normalize to 0-255 and save as uint8_t I then normalize and and display And what shows up: It is maddening

If I subtract 2 StereoDepth frames from each other how to output in OpenCV

AdamPolak

I created a NN that subtracts 2 depth frames:

class DiffImgs(nn.Module):

def forward(self, img1, img2):

if img1.shape != (1, 1, 960, 1632) or img2.shape != (1, 1, 960, 1632):

raise ValueError('Input images must have shape (1, 1, 960, 1632)')

return torch.sub(img1, img2)

I get the output as a vector<float> by using:

auto depth_diff_message = queues["depth_diff"]->get<dai::NNData>();

std::vector<float> depth_diff_data = depth_diff_message->getFirstLayerFp16()

The input of StereoDepth with subpixel to the model is UINT16
The model does the diff and outputs at Fp16
Because I subtracted it means the results can be negative or positive
getFirstLayerFp16 outputs a 1D vector of floats
In the depthai C++ repo it looks like there needs to be some conversion from NNData to get into a CV frame:
https://github.com/luxonis/depthai-core/blob/main/examples/utility/utility.cpp
cv::Mat fromPlanarFp16(const std::vector<float>& data, int w, int h, float mean, float scale){

cv::Mat frame = cv::Mat(h, w, CV_8UC3);

for(int i = 0; i < w*h; i++) {

auto b = data.data()[i + w*h * 0] * scale + mean;

frame.data[i*3+0] = (uint8_t)b;

}

for(int i = 0; i < w*h; i++) {

auto g = data.data()[i + w*h * 1] * scale + mean;

frame.data[i*3+1] = (uint8_t)g;

}

for(int i = 0; i < w*h; i++) {

auto r = data.data()[i + w*h * 2] * scale + mean;

frame.data[i*3+2] = (uint8_t)r;

}

return frame;

}

What is the proper way to convert a diff'd StereoDepth frame from vector<float> into cv::Mat?

erik

Hi AdamPolak ,
As it's INT16, you'd need to implement something similar to depth -> pointcloud conversion docs for your "model" as well:
https://docs.luxonis.com/en/latest/pages/tutorials/device-pointcloud/#on-device-pointcloud-nn-model
Thanks, Erik

AdamPolak

erik Huge help thank you. Will dig into this.

AdamPolak

erik

I have updated the model to take in the conversion of a depth map from U8 to Fp16

I can't figure out how to "get back" to the depth frame from the resulting NNData.

1. If it is a vector of floats that I pull out of NNData that are fp16, does that mean I need to reduce it again to U8? How is the conversation made so I know how to undo it.

If I set sub-pixel = true, the data type changes to UINT16, how would that change what I need to do?

erik

Hi AdamPolak ,

I don't think that's possible - afaik NN can only output FP16.
It's the same, you are expecting depth (INT16) anyways, not disparity (where it changes from INT8 to INT16 when subpixel is enabled).

AdamPolak

erik

I am close, I am getting depth values that make sense in fp16 from the model.

The issue is I do not know how to convert those floats back into a cv::mat

I tried adding the floats to a cv::Mat.data() just by going 1 by 1 but it seems to be encoded differently. How does the 1D vector of floats compare to a 2D 1 channel image? Should I be alternating every float or something for height/width?

This seems to be my last step.

erik

AdamPolak please check how it's done on other demos:
https://github.com/luxonis/depthai-experiments/blob/master/gen2-custom-models/concat.py#L58

inNn = np.array(qNn.get().getData())
frame = inNn.view(np.float16).reshape(shape).transpose(1, 2, 0).astype(np.uint8)
cv2.imshow("Concat", frame)

AdamPolak

erik

I have been staring at that example so hard I think I have it memorized lol.

The example uses 3 channels and uses numpy to reshape. The C++ version uses this utility function to shape the 3 channels:
https://github.com/luxonis/depthai-core/blob/main/examples/utility/utility.cpp
https://github.com/luxonis/depthai-core/blob/main/examples/NeuralNetwork/concat_multi_input.cpp

So it doesn't have to deal with translating it from 0-65535 to 0-255 In a way that can be displayed.
Interpreting it as a CV32FC1 frame and then using this approach:
https://github.com/luxonis/depthai-core/blob/125feb8c2e16ee4bf71b7873a7b990f1c5f17b18/examples/StereoDepth/depth_preview.cpp#LL54C43-L54C43
frame.convertTo(frame, CV_8UC1, 255 / depth->initialConfig.getMaxDisparity());

Leaves it scrambled as well.

I can't figure what the heck I am doing wrong.

erik

AdamPolak have you consulted GPT4 already? Usually it's fairly smart about such python->cpp conversions.

AdamPolak

erik

Like you wouldn't believe. It seems like since it got nerfed in the latest update it can't do hardcore things anymore. It got the order of how pytorch interprets columns/rows wrong and threw me off for half a day lol.

AdamPolak

erik

I subtract the depth frame from 0's in the model:

So it is just outputting the same depth frame values.
I get the the NNData from the queue and check that the size is the same as the expected 1632x960 and it is
I create a cv::Mat and then iterate through the floats, normalize to 0-255 and save as uint8_t
I then normalize and and display

And what shows up:

It is maddening

erik

AdamPolak does it work with python (numpy)? If not, I can help with that. I am not familiar with cpp though, but I'd guess it shouldn't be too hard to implement the same logic.

AdamPolak

erik

Good call, I will try in python and see what is up.

AdamPolak

Rewrote it in python and now there is a bit of an actual outline of me in my seat in the middle of what is showing up

But I realize that I was reshaping it the wrong way and changed it to reshape how it was in the example and got it scrambled:

Now I am trying to figure out why when I messed up the reshaping was it the closest it was to looking accurate

AdamPolak

Why the heck can I make out an image barely of my hand when I scramble the reshape

erik

AdamPolak feel free to post source code (not screenshots) and perhaps we can take a look into it as well.

AdamPolak

@erik This is the depthai code:

import numpy as np
import cv2
import depthai as dai


resolution = (1632,960) # 24 FPS (without visualization)
lrcheck = True  # Better handling for occlusions
extended = False  # Closer-in minimum depth, disparity range is doubled
subpixel = True  # True  # Better accuracy for longer distance, fractional disparity 32-levels

p = dai.Pipeline()

# Configure Mono Camera Properties
left = p.createMonoCamera()
left.setResolution(dai.MonoCameraProperties.SensorResolution.THE_800_P)
left.setBoardSocket(dai.CameraBoardSocket.LEFT)

right = p.createMonoCamera()
right.setResolution(dai.MonoCameraProperties.SensorResolution.THE_800_P)
right.setBoardSocket(dai.CameraBoardSocket.RIGHT)

stereo = p.createStereoDepth()
left.out.link(stereo.left)
right.out.link(stereo.right)

# Set stereo depth options
stereo.setDefaultProfilePreset(dai.node.StereoDepth.PresetMode.HIGH_DENSITY)
config = stereo.initialConfig.get()
config.postProcessing.speckleFilter.enable = True
config.postProcessing.speckleFilter.speckleRange = 60
config.postProcessing.temporalFilter.enable = True

config.postProcessing.spatialFilter.holeFillingRadius = 2
config.postProcessing.spatialFilter.numIterations = 1
config.postProcessing.thresholdFilter.minRange = 700  # mm
config.postProcessing.thresholdFilter.maxRange = 7000  # mm
config.censusTransform.enableMeanMode = True
config.costMatching.linearEquationParameters.alpha = 0
config.costMatching.linearEquationParameters.beta = 2
stereo.initialConfig.set(config)
stereo.setLeftRightCheck(lrcheck)
stereo.setExtendedDisparity(extended)
stereo.setSubpixel(subpixel)
stereo.setDepthAlign(dai.CameraBoardSocket.RGB)
stereo.setRectifyEdgeFillColor(0)  # Black, to better see the cutout

# Depth -> Depth Diff
nn = p.createNeuralNetwork()
nn.setBlobPath("diff_images_simplified_openvino_2021.4_4shave.blob")
stereo.disparity.link(nn.inputs["input1"])

depthDiffOut = p.createXLinkOut()
depthDiffOut.setStreamName("depth_diff")
nn.out.link(depthDiffOut.input)

with dai.Device(p) as device:
    qDepthDiff = device.getOutputQueue(name="depth_diff", maxSize=4, blocking=False)
    while True:
        depthDiff = qDepthDiff.get()

        # Shape it here
        floatVector = depthDiff.getFirstLayerFp16()
        diff = np.array(floatVector).reshape(resolution[0], resolution[1])
        
        colorize = cv2.normalize(diff, None, 255, 0, cv2.NORM_INF, cv2.CV_8UC1)
        cv2.applyColorMap(colorize, cv2.COLORMAP_JET)
        cv2.imshow("Diff", colorize)
        if cv2.waitKey(1) == ord('q'):
                    break

This is the pytorch code. I no longer subtract from dummy frame and just pass through depth:

#! /usr/bin/env python3

from pathlib import Path
import torch
from torch import nn
import blobconverter
import onnx
from onnxsim import simplify
import sys

# Define the model
class DiffImgs(nn.Module):
    def forward(self, depth):
        # We will be inputting UINT16 but interprets as UINT8
        # So we need to adjust to account of the 8 bit shift
        depthFP16 = 256.0 * depth[:,:,:,1::2] + depth[:,:,:,::2]
        return depthFP16
        # depthFP16 = depthFP16.view(1, -1)
        # depthFP16_shape = depthFP16.shape
        # Create a dummy frame 0 frame to test if depth can be recaptured on host
        # dumy_frame = torch.zeros(depthFP16_shape, dtype=torch.float16)
        # return torch.sub(depthFP16, dumy_frame)

# Instantiate the model
model = DiffImgs()

# Create dummy input for the ONNX export
input1 = torch.randn(1, 1, 960, 1632 * 2, dtype=torch.float16)
input2 = torch.randn(1, 1, 960, 1632 * 2, dtype=torch.float16)

onnx_file = "diff_images.onnx"

# Export the model
torch.onnx.export(model,               # model being run
                  (input1),    # model input (or a tuple for multiple inputs)
                  onnx_file,        # where to save the model (can be a file or file-like object)
                  opset_version=12,    # the ONNX version to export the model to
                  do_constant_folding=True,  # whether to execute constant folding for optimization
                  input_names = ['input1'],   # the model's input names
                  output_names = ['output'])

# Simplify the model
onnx_model = onnx.load(onnx_file)
onnx_simplified, check = simplify(onnx_file)
onnx.save(onnx_simplified, "diff_images_simplified.onnx")

# Use blobconverter to convert onnx->IR->blob
blobconverter.from_onnx(
    model="diff_images_simplified.onnx",
    data_type="FP16",
    shaves=4,
    use_cache=False,
    output_dir="../",
    optimizer_params=[],
    compile_params=['-ip U8'],    
)

erik

AdamPolak why / how did you get resolution of (1632,960)?

jeremie_m

AdamPolak

Thank you Adam, I have questions about the model 'diff_images_simplified_openvino_2021.4_4shave.blob', is it generated by the pytorch code here?

AdamPolak This is the pytorch code.

Is it still using the dummy input or the depth input here is the result of the subtraction?

AdamPolak def forward(self, depth):

jeremie_m

AdamPolak depthFP16 = 256.0 * depth[:,:,:,1::2] + depth[:,:,:,::2]

Is the subtraction executed here?

jeremie_m

AdamPolak quite a lot of depth filters enabled to make this diff work

Thanks, Adam, 9-11 FPS maybe enough for me, I have to try to make the filters work in the host if the rate is too low.

Is the config of depth filters is set as you mentioned in the code here or it's more complex than the parameters here?

AdamPolak This is the depthai code

AdamPolak

erik

That is the resolution of my camRgb preview for person detection that I got from this example:https://github.com/luxonis/depthai-experiments/blob/30e2460557a3209770eb8943db41bc997a423212/gen2-pedestrian-reidentification/api/main_api.py#L22

I think I found what you found. Even if you set a depth preview a certain size, it won't make the image that big. I see now that the biggest that comes out is 1920,1080.

If my camera preview is 1632x960, how can I rgb align so I can find the spatial location of each RGB pixel?

I am having trouble understanding how depthpreview, depth resolution, rgb align, and rgb resolution all play together.

I had at one point added a ValueError for the ML model if the input size was not the same as expected, but apparently that doesn't trigger anything in OpenVINO.

AdamPolak

erik

If I have an rgb at 1632 x 960 preview and 1080p resolution

Does this mean if I do depth align it listens to the resolution not the preview? How can I use this for spatial location calculations, would I have to resize the depth frame?

erik

AdamPolak yes it aligns with resolution, not the preview. spatial loc calc knows image transformation (all crops/resizes/etc) so it will map to correct depth region.