manurare / 360monodepth

Code release for 360monodepth. With our framework we achieve monocular depth estimation for high resolution 360° images based on aligning and blending perspective depth maps.
https://manurare.github.io/360monodepth/
MIT License
159 stars 33 forks source link

Crash when transposing A matrix #18

Closed patriciogonzalezvivo closed 7 months ago

patriciogonzalezvivo commented 11 months ago

Hi, the main.py python script crashes at this line part of compute_linear_system_matrices() as part of the poisson blending process.

    self.eigen_solver.A = self.A.transpose().dot(self.A)

Any idea how should I debug or solve it?

patriciogonzalezvivo commented 11 months ago

Breaking this up in steps, seams crash happens on the dot product

        blocks.append(self.fidelity_weight * scipy.sparse.eye(blocks[0].shape[1]))
        mat_A = self.concatenate_csr_matrices_by_row(blocks)
        self.A = mat_A
        if self.eigen_solver is not None:
            print("A_ = self.A")
            A_ = self.A
            print("self.A.transpose().dot(A_)")
            B_ = A_.transpose()
            print("self.dot")
            C_ = B_.dot(self.A)
            print("Assign it to self.eigen_solver.A")
            self.eigen_solver.A = C_

        # self.x_grad_mat = x_grad_mat
        # self.y_grad_mat = y_grad_mat
        print("Done computing matrices")

Produce

I1228 12:35:09.807940  6531 depthmap_stitcher.cpp:233] Output the aligned coefficients to file :data/debug/data/debug/depthmapAlignPy_coeff.json
return alignment coefficient and depth maps...
2023-12-28 12:35:09,821 - __main__ - INFO - 4) blender to ERP image
2023-12-28 12:35:09,821 - utility.blending - DEBUG - stitch the tangent image 0
2023-12-28 12:35:09,939 - utility.blending - DEBUG - stitch the tangent image 1
2023-12-28 12:35:10,051 - utility.blending - DEBUG - stitch the tangent image 2
2023-12-28 12:35:10,162 - utility.blending - DEBUG - stitch the tangent image 3
2023-12-28 12:35:10,271 - utility.blending - DEBUG - stitch the tangent image 4
2023-12-28 12:35:10,381 - utility.blending - DEBUG - stitch the tangent image 5
2023-12-28 12:35:10,417 - utility.blending - DEBUG - stitch the tangent image 6
2023-12-28 12:35:10,452 - utility.blending - DEBUG - stitch the tangent image 7
2023-12-28 12:35:10,488 - utility.blending - DEBUG - stitch the tangent image 8
2023-12-28 12:35:10,521 - utility.blending - DEBUG - stitch the tangent image 9
2023-12-28 12:35:10,557 - utility.blending - DEBUG - stitch the tangent image 10
2023-12-28 12:35:10,593 - utility.blending - DEBUG - stitch the tangent image 11
2023-12-28 12:35:10,629 - utility.blending - DEBUG - stitch the tangent image 12
2023-12-28 12:35:10,667 - utility.blending - DEBUG - stitch the tangent image 13
2023-12-28 12:35:10,702 - utility.blending - DEBUG - stitch the tangent image 14
2023-12-28 12:35:10,739 - utility.blending - DEBUG - stitch the tangent image 15
2023-12-28 12:35:10,856 - utility.blending - DEBUG - stitch the tangent image 16
2023-12-28 12:35:10,969 - utility.blending - DEBUG - stitch the tangent image 17
2023-12-28 12:35:11,078 - utility.blending - DEBUG - stitch the tangent image 18
2023-12-28 12:35:11,191 - utility.blending - DEBUG - stitch the tangent image 19
A_ = self.A
self.A.transpose().dot(A_)
self.dot
[1]    6531 killed     python main.py -i data/0001.jpg --persp_monodepth=midas2 

Any ideas what could be happening here?

manurare commented 11 months ago

Hello,

It is weird that it doesn't throw anything more descriptive. My guess is that you are running out of RAM. For example if I run this code in my laptop it also gets killed:

A = scipy.sparse.diags(np.arange(100000000))
A.transpose().dot(A)

Of course this matrix would belong to an image of 100 megapixels. But it only stores 1 float per row. I am not sure about the sizes of your images, but the A matrix for Poisson blending stores more than 1 float per row.

patriciogonzalezvivo commented 11 months ago

I'm using this image which is the same provided by on the original repo as an example

2023-12-29 12:50:18,502 - utility.blending - DEBUG - stitch the tangent image 18
2023-12-29 12:50:18,628 - utility.blending - DEBUG - stitch the tangent image 19
self.A (85983232, 2097152)
transpose()
(2097152, 85983232)
dot()
[1]    6341 killed     python main.py -i data/0001.jpg --persp_monodepth=midas2 

The crash happens in at the dot product. A matrix before that operation is (85983232, 2097152). Do you mind telling me what is this transposition and dot product about?

manurare commented 11 months ago

Hi. For the solver you need a square A matrix. That's why you need to multiply A by its transpose