Segment fusion based robust semantic segmentation of scenes using a graph clusterer

1 . A computing system comprising:

a network controller to obtain data corresponding to a scene;

a processor coupled to the network controller; and

a memory including a set of instructions, which when executed by the processor, cause the processor to:

use a first neural network to generate semantic features,

use a second neural network to generate instance features,

identify a plurality of segments using a graph clusterer based on the semantic features, the instance features, and point cloud data associated with the scene,

fuse the plurality of segments into a plurality of instances, wherein the plurality of segments is to be fused into the plurality of instances via a learnable self-attention based network, and

select classification labels for the plurality of instances.

2 . The computing system of claim 1 , wherein the computing system is end-to-end learnable.

3 . The computing system of claim 1 , wherein the plurality of segments is to be fused into the plurality of instances based on an instance loss function, a segment loss function, and a distance margin parameter.

4 . The computing system of claim 3 , wherein the segment loss function is to penalize fusion mispredictions and separation mispredictions.

5 . The computing system of claim 1 , wherein to select the classification labels, the instructions, when executed, further cause the processor to:

generate, on a per instance basis, a semantic label for each voxel in the instance, and

select the classification label based on semantic labels of voxels in the instance.

6 . The computing system of claim 1 , wherein the plurality of segments is to be variable in size.

7 . At least one non-transitory computer readable storage medium comprising a set of instructions, which when executed by a computing system, cause the computing system to:

use a first neural network to generate semantic features;

use a second neural network to generate instance features;

identify a plurality of segments using a graph cluster based on the semantic features, the instance features, and point cloud data associated with a scene;

fuse the plurality of segments into a plurality of instances, wherein the plurality of segments is to be fused into the plurality of instances via a learnable self-attention based network; and

select classification labels for the plurality of instances.

8 . The at least one non-transitory computer readable storage medium of claim 7 , wherein the plurality of segments is to be fused into the plurality of instances based on an instance loss function, a segment loss function, and a distance margin parameter.

9 . The at least one non-transitory computer readable storage medium of claim 8 , wherein the segment loss function is to penalize fusion mispredictions and separation mispredictions.

10 . The at least one non-transitory computer readable storage medium of claim 7 , wherein to select the classification labels, the instructions, when executed, further cause the computing system to:

generate, on a per instance basis, a semantic label for each voxel in the instance; and

select the classification label based on semantic labels of voxels in the instance.

11 . The at least one non-transitory computer readable storage medium of claim 7 , wherein the plurality of segments is to be variable in size.

12 . A semiconductor apparatus comprising:

one or more substrates; and

logic coupled to the one or more substrates, wherein the logic is implemented at least partly in one or more of configurable or fixed-functionality hardware, the logic to:

use a first neural network to generate semantic features,

use a second neural network to generate instance features,

identify a plurality of segments using a graph clusterer based on the semantic features, the instance features, and point cloud data associated with a scene,

fuse the plurality of segments into a plurality of instances, wherein the plurality of segments is to be fused into the plurality of instances via a learnable self-attention based network, and

select classification labels for the plurality of instances.

13 . The semiconductor apparatus of claim 12 , wherein the plurality of segments is to be fused into the plurality of instances based on an instance loss function, a segment loss function, and a distance margin parameter.

14 . The semiconductor apparatus of claim 13 , wherein the segment loss function is to penalize fusion mispredictions and separation mispredictions.

15 . The semiconductor apparatus of claim 12 , wherein to select the classification labels, the logic is to:

generate, on a per instance basis, a semantic label for each voxel in the instance; and

select the classification label based on semantic labels of voxels in the instance.

16 . The semiconductor apparatus of claim 12 , wherein the plurality of segments is to be variable in size.

17 . The semiconductor apparatus of claim 12 , wherein the logic coupled to the one or more substrates includes transistor channel regions that are positioned within the one or more substrates.

18 . A method comprising:

using a first neural network to generate semantic features,

using a second neural network to generate instance features,

identifying a plurality of segments using a graph cluster based on the semantic features, the instance features, and point cloud data associated with a scene;

fusing the plurality of segments into a plurality of instances, wherein the plurality of segments is fused into the plurality of instances via a learnable self-attention based network; and

selecting classification labels for the plurality of instances.

19 . The method of claim 18 , wherein the plurality of segments is fused into the plurality of instances based on an instance loss function, a segment loss function, and a distance margin parameter.

20 . The method of claim 19 , wherein the segment loss function penalizes fusion mispredictions and separation mispredictions.

21 . The method of claim 18 , wherein selecting the classification labels includes:

generating, on a per instance basis, a semantic label for each voxel in the instance; and

selecting the classification label based on semantic labels of voxels in the instance.

22 . The method of claim 18 , wherein the plurality of segments is variable in size.