Method for generating training data, image semantic segmentation method and electronic device

1. A method for generating training data, comprising:

according to indexes of the object models, defining a corresponding category tag for an object model in a three-dimensional scenario, wherein each category tag represents the category of the object model;

acquiring a plurality of corresponding scenario images by modifying scenario parameters of the three-dimensional scenario, wherein the scenario parameters are some factors affecting the appearance performance of the three-dimensional model, and the scenario images are planar images of the three-dimensional scenario that are acquired under the effect of different factors, and wherein the scenario images corresponds to the scenario parameters;

rendering the object model to a monochromic material corresponding to the category tag of the object model;

acquiring a semantic segmentation image corresponding to each scenario image based on the rendered object model; and

storing each scenario image and the semantic segmentation image corresponding to the scenario image as a set of training data;

wherein the rendering the object model to the monochromic material corresponding to the category tag of the object model comprises:

judging whether a transparent object model achieves a shielding effect against the object model thereafter;

when the transparent object model achieves the shielding effect, assigning a category tag for the transparent object model and rendering the transparent object model to the corresponding monochromic material; and

when the transparent object model fails to achieve the shielding effect, maintaining the transparent object model as transparent in the semantic segmentation image or deleting the transparent object model from the semantic segmentation image.

2. The method according to claim 1 , wherein the acquiring the plurality of scenario images corresponding to the three-dimensional scenario comprises:

dynamically adjusting illumination conditions of the three-dimensional scenario; and

capturing scenario images under different illumination conditions by a virtual camera.

3. The method according to claim 1 , wherein the acquiring the plurality of corresponding scenario images by modifying the scenario parameters of the three-dimensional scenario comprises:

defining a photographing trajectory of the virtual camera; and

capturing scenario images from different view angles when the virtual camera moves along the photographing trajectory.

4. The method according to claim 1 , wherein the acquiring the semantic segmentation image corresponding to each scenario image based on the rendered object model comprises:

acquiring the semantic segmentation image corresponding to each scenario image based on the rendered object model when illumination of the three-dimensional scenario is turned off.

5. The method according to claim 1 , wherein the rendering the object model to the monochromic material corresponding to the category tag of the object model comprises:

determining a depth sequence of the object model in the scenario image; and

rendering in sequence the object model to the monochromic material corresponding to the category tag of the object model.

6. An image semantic segmentation method, comprising:

according to indexes of the object models, defining a corresponding category tag for an object model in a three-dimensional scenario, wherein each category tag represents the category of the object model;

acquiring a plurality of corresponding scenario images by modifying scenario parameters of the three-dimensional scenario, wherein the scenario parameters are some factors affecting the appearance performance of the three-dimensional model, and the scenario images are planar images of the three-dimensional scenario that are acquired under the effect of different factors, and wherein the scenario images corresponds to the scenario parameters;

rendering the object model to a monochromic material corresponding to the category tag of the object model;

acquiring a semantic segmentation image corresponding to each scenario image based on the rendered object model; and

storing each scenario image and the semantic segmentation image corresponding to the scenario image as a set of training data;

using the scenario image and the corresponding semantic segmentation image as training data;

wherein the rendering the object model to the monochromic material corresponding to the category tag of the object model comprises:

judging whether a transparent object model achieves a shielding effect against the object model thereafter;

when the transparent object model achieves the shielding effect, assigning a category tag for the transparent object model and rendering the transparent object model to the corresponding monochromic material; and

when the transparent object model fails to achieve the shielding effect, maintaining the transparent object model as transparent in the semantic segmentation image or deleting the transparent object model from the semantic segmentation image.

7. The image semantic segmentation method according to claim 6 , wherein the acquiring the plurality of scenario images corresponding to the three-dimensional scenario comprises:

dynamically adjusting illumination conditions of the three-dimensional scenario; and

capturing scenario images under different illumination conditions by a virtual camera.

8. The image semantic segmentation method according to claim 6 , wherein the acquiring the plurality of corresponding scenario images by modifying the scenario parameters of the three-dimensional scenario comprises:

defining a photographing trajectory of the virtual camera; and

capturing scenario images from different view angles when the virtual camera moves along the photographing trajectory.

9. The image semantic segmentation method according to claim 6 , wherein the acquiring the semantic segmentation image corresponding to each scenario image based on the rendered object model comprises:

acquiring the semantic segmentation image corresponding to each scenario image based on the rendered object model when illumination of the three-dimensional scenario is turned off.

10. The image semantic segmentation method according to claim 9 , wherein the rendering the object model to the monochromic material corresponding to the category tag of the object model comprises:

determining a depth sequence of the object model in the scenario image; and

rendering in sequence the object model to the monochromic material corresponding to the category tag of the object model.

11. An electronic device, comprising:

at least one processor; and

a memory communicably connected to the at least one processor; wherein

the memory stores an instruction program executable by the at least one processor, wherein, the instruction program, when being executed by the at least one processor, cause the at least one processor to perform a method comprising:

according to indexes of the object models, defining a corresponding category tag for an object model in a three-dimensional scenario, wherein each category tag represents the category of the object model;

acquiring a plurality of corresponding scenario images by modifying scenario parameters of the three-dimensional scenario, wherein the scenario parameters are some factors affecting the appearance performance of the three-dimensional model, and the scenario images are planar images of the three-dimensional scenario that are acquired under the effect of different factors, and wherein the scenario images corresponds to the scenario parameters;

rendering the object model to a monochromic material corresponding to the category tag of the object model;

acquiring a semantic segmentation image corresponding to each scenario image based on the rendered object model; and

storing each scenario image and the semantic segmentation image corresponding to the scenario image as a set of training data;

wherein the rendering the object model to the monochromic material corresponding to the category tag of the object model comprises:

judging whether a transparent object model achieves a shielding effect against the object model thereafter;

when the transparent object model achieves the shielding effect, assigning a category tag for the transparent object model and rendering the transparent object model to the corresponding monochromic material; and

when the transparent object model fails to achieve the shielding effect, maintaining the transparent object model as transparent in the semantic segmentation image or deleting the transparent object model from the semantic segmentation image.

12. The electronic device according to claim 11 ,

wherein the acquiring the plurality of scenario images corresponding to the three-dimensional scenario comprises:

dynamically adjusting illumination conditions of the three-dimensional scenario; and

capturing scenario images under different illumination conditions by a virtual camera.

13. The electronic device according to claim 11 ,

wherein the acquiring the plurality of corresponding scenario images by modifying the scenario parameters of the three-dimensional scenario comprises:

defining a photographing trajectory of the virtual camera; and

capturing scenario images from different view angles when the virtual camera moves along the photographing trajectory.

14. The electronic device according to claim 11 ,

wherein the acquiring the semantic segmentation image corresponding to each scenario image based on the rendered object model comprises:

acquiring the semantic segmentation image corresponding to each scenario image based on the rendered object model when illumination of the three-dimensional scenario is turned off.

15. The electronic device according to claim 11 ,

wherein the rendering the object model to the monochromic material corresponding to the category tag of the object model comprises:

determining a depth sequence of the object model in the scenario image; and

rendering in sequence the object model to the monochromic material corresponding to the category tag of the object model.