The application discloses an outdoor scene three-dimensional
reconstruction method based on a
deep learning technology. The method comprises the following steps: step 1, inputting pictures taken from multiple perspectives to generate a
voxel grid; step 2, sending the
voxel grid into a
scaffold, and outputting multi-scale feature fields through 3D sparse
convolution, feature
pyramid and attention mechanism; step 3, using 3D CNN to perform geometric decoding on the multi-scale features to output parameterized ellipsoids; step 4, performing post-
processing on the parameterized ellipsoids to realize
ellipsoid instantiation, performing differentiable rasterization on
Gaussian ellipsoids, and performing filtering operation on the rendered images after rasterization; and step 5, comparing the filtered images with real images, combining light loss and image loss to perform back propagation and update parameters. Through the design of a new
scaffold, a filtering module and a
loss function, the application realizes the fusion of multi-scale features and the efficient
processing of sparse data, the model can better filter out
noise, provide clearer and more consistent edge and texture information, solve the problems of
noise, geometric edge sawtooth and artifacts existing in the rasterization process, enhance the contrast of different light regions, solve the problem that different materials and textures in an outdoor scene perform differently under
light irradiation, and thus improve the reconstruction accuracy.