OpenCV----Adabins单目深度估计LibTorch推理

题目要求：学习了解单目深度估计模型Adabins，根据python源码集成到现有OpenCV系列模型中。
AdaBins 论文：Adabins: Depth Estimation using adaptive bins
AdaBins 源码：AdaBins GitHub

分析：
1）了解Adabins的基本原理和代码理解
2）将模型转化为更加方便高效的LibTorch模型并在opencv中完成推理过程（并验证）

• 结果展示：
  
• Pytorch转LibTorch

	import onnx
	import onnxruntime as ort
	import cv2
	
	import matplotlib.pyplot as plt
	
	def torch2LibTorch():
	    MIN_DEPTH = 1e-3
	    MAX_DEPTH_NYU = 10
	 
	    N_BINS = 256
	 
	    pretrained_path = "./pretrained/AdaBins_nyu.pt"
	 
	    # device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
	    device = 'cpu'
	 
	    # load model
	    model = UnetAdaptiveBins.build(n_bins=N_BINS, min_val=MIN_DEPTH, max_val=MAX_DEPTH_NYU)
	    model, _, _ = model_io.load_checkpoint(pretrained_path, model)
	 
	    # set the model to inference mode
	    model.to(device)
	    model.eval()
	
	    ## convert to trace model
	    example = torch.rand(1, 3, 480, 640).to(device)
	    traced_script_module = torch.jit.trace(model, example)
	    traced_script_module.save('adabins_trace.pt')

• OpenCV和LibTorch模型推理

	class adabins
	{
	public:
		adabins(int h, int w, const string& model_file = "model/adabins_trace.pt") {
			this->inHeight = h;
			this->inWidth = w;
			net = torch::jit::load(model_file);
		};
		Mat depth(const Mat& frame);
	private:
		torch::jit::script::Module net;
		int inWidth;
		int inHeight;
	};
	
	Mat adabins::depth(const Mat& frame){
		torch::NoGradGuard no_grad;
		Mat img = frame.clone();
	    int img_h = img.rows;
	    int img_w = img.cols;
	    resize(img, img, Size(this->inHeight, this->inWidth));
	    torch::Tensor tensor_image = torch::from_blob(img.data, { 1, img.rows, img.cols,3 }, torch::kU8).to(torch::kFloat)/255.0;
	    tensor_image = tensor_image.permute({0,3,1,2});
	    tensor_image = tensor_image.to(at::kCPU);
	
	    net.to(at::kCPU);
	    net.eval();
	    torch::Tensor out_tensor = net.forward({tensor_image}).toTuple()->elements()[1].toTensor();
	    out_tensor = out_tensor.detach().permute({0,2,3,1}).squeeze();
		auto min_t = 0.1;
		auto max_t = 10;
	    out_tensor = out_tensor.clamp(min_t, max_t);
	    out_tensor = (out_tensor - min_t) / (max_t - min_t);
	    out_tensor = out_tensor.mul(255.0).clamp(0, 255).to(torch::kU8);
	    out_tensor = out_tensor.to(at::kCPU);
	    Mat resultImg(320, 240, CV_8UC1, out_tensor.data_ptr()); 
	    resize(resultImg, resultImg,Size(img_w, img_h));
		// imwrite("inference/test.png", resultImg);
	    applyColorMap(resultImg, resultImg, COLORMAP_JET);
		return resultImg;
	}

// module test
// int main()
// {   
//     Mat frame = imread("inference/classroom.jpg", 1);
// 	if (frame.empty()) {
// 		printf("could not load image...\n");
// 		return -1;
// 	}
// 	int h = 480, w = 640;
// 	adabins net(h, w);
// 	Mat depth = net.depth(frame);
// 	Mat res = viewer({frame, depth});
// 	static const string kWinName = "Deep learning object detection in OpenCV";
// 	namedWindow(kWinName, WINDOW_NORMAL);
// 	imshow(kWinName, res);
// 	waitKey(0);
// 	destroyAllWindows();
// 	return 0;
// }

• 集成到系统框架

else if(config.model_name == "adabins"){
		int h = 480, w = 640;
		adabins net(h, w);
		Mat depth = net.depth(srcimg);
		Mat res = viewer({srcimg, depth});
		kWinName = "Deep learning depth estimation ADABINS in OpenCV";
		namedWindow(kWinName, WINDOW_NORMAL);
		imshow(kWinName, res);
		waitKey(0);
		destroyAllWindows();

• 小结：
  1. LibTroch转换模型时报错：numpy_T算子不支持
     解决方法：
     embeddings = embeddings + self.positional_encodings[:embeddings.shape[2], :].transpose(1, 0).unsqueeze(0)
  2. CMakeLists.txt部分变化

  add_library(adabins adabins.cpp)
  target_link_libraries(adabins opencv_world460 ${TORCH_LIBRARIES})