Request Multi-Class Support + Export Script for YOLO-Pose

[Nasaravira20]

Summary

DeepStream’s current YOLO-Pose parser supports only single-class pose models.
However, modern Ultralytics YOLO-Pose models (YOLOv8/YOLO11) fully support
multi-class pose detection.

I exported my detector model to ONNX and verified that the export is correct:

• Input shape: ['batch', 3, 640, 640]
• Number of outputs: 3
• Output 0 → (1, 8400, 4)
• Output 1 → (1, 8400, 1)
• Output 2 → (1, 8400, 1)

Since pose models output one confidence per anchor, the ONNX is valid
the issue is with the DeepStream parsing logic.

Problem

The DeepStream YOLO-Pose parser:

• Assumes all detections belong to class 0
• Does not support num_classes > 1
• Cannot map labels or generate metadata for multiple classes

As a result, any custom YOLO-Pose model trained with multiple classes
cannot be used in DeepStream.

Expected Features

1. Multi-class parsing logic, similar to YOLO-Detect:

   • Ability to assign class IDs per detection
   • Per-class filtering
   • Label mapping via .txt or .yaml
   • Proper nvinfer metadata output

2. Export script for YOLO-Pose:

   • A DeepStream-compatible ONNX/TensorRT export utility
   • Matching output formatting for multi-class pose
   • Config parameters for number of keypoints, number of classes

Why This Matters

Multi-class pose models are widely used in:

• Industrial safety (PPE detection + pose)
• Retail analytics (customer + product interaction)
• Robotics (multiple object types + pose)
• Sports analytics

DeepStream currently lacks official multi-class pose support, limiting adoption.

Request

Please add:

• Multi-class YOLO-Pose parsing support
• Official export script for custom multi-class YOLO-Pose models