Feature: Add point-cloud semantic segmentation task and PointNet model family

Request type

New model or checkpoint support

Problem and use case

TensorRT-Model-Connect currently provides model recipes for several vision tasks, including image segmentation, but does not appear to provide an end-to-end task contract or model recipe for semantic segmentation of 3D point clouds.

Point-cloud semantic segmentation is commonly used in robotics, autonomous systems, indoor scene understanding, and other 3D perception workloads. In these applications, the model receives a set of 3D points (optionally with additional per-point features such as color) and predicts a semantic class for each point.

Today, users who want to deploy a point-cloud segmentation model cannot use the existing image-segmentation path directly because the input/output contract is different: image segmentation operates on a dense image grid, while point-cloud segmentation operates on an unordered set of 3D points and produces per-point predictions.

It would be useful for TensorRT-Model-Connect to support point-cloud semantic segmentation as an end-to-end workflow, with PointNet as an initial model family/reference implementation.

Proposed outcome

Add end-to-end support for point-cloud semantic segmentation, using PointNet as the initial supported model family.

At a high level, the desired user workflow would be:

• Build a supported PointNet semantic-segmentation checkpoint into a TensorRT-Model-Connect bundle.
• Provide point-cloud data as inference input.
• Run inference through the native TensorRT-Model-Connect runtime.
• Return per-point semantic predictions.
• Add end-to-end validation against a reference implementation.

For an initial version, I think it would be reasonable to keep the scope small and focus on a basic PointNet semantic-segmentation model before considering more complex point-cloud architectures such as PointNet++, sparse-convolution networks, or instance segmentation.

I would be interested in contributing the initial implementation and would appreciate feedback on the preferred task/runtime contract before starting the larger changes.

Model and target details

Initial model family:

• PointNet
• Task: 3D point-cloud semantic segmentation
• Input: a point set with XYZ coordinates and optionally additional per-point features
• Output: per-point semantic class predictions

Initial target:

• NVIDIA GPU
• TensorRT native runtime
• FP16 as an initial precision target

I am still evaluating the most appropriate publicly available PointNet semantic-segmentation checkpoint for the E2E reference case and would be happy to align with the maintainers on the preferred checkpoint and dataset.

Alternatives considered

No response

Additional context

The main goal of this request is not to introduce a large 3D perception stack in one change, but to establish the smallest useful end-to-end point-cloud segmentation path first.

If the direction is acceptable, I can start by examining the closest existing model-family and E2E implementations and prepare a small prototype before submitting the full contribution.

Submission checks

• I searched open and closed issues and found no duplicate request.
• I removed secrets, private/internal evidence, personal paths, and restricted artifacts.

No specific files are named. Start by examining the closest existing model-family and end-to-end implementations, then align with maintainers on the task/runtime contract, checkpoint, and dataset. Done means a basic PointNet semantic-segmentation workflow builds a bundle, accepts XYZ point-cloud input, runs through the native TensorRT runtime, returns per-point predictions, and has end-to-end validation.