🟠 Object Detection (Recognize Orientation)

Function Description

The operator performs object detection on input images based on deep learning. Unlike standard rectangular box detection, this operator can be used to recognize and output object rotation angles, thus distinguishing target orientations and generating rotated rectangle boxes that fit objects more tightly.

Usage Scenarios

When application scenarios require not only locating objects but also knowing their precise placement pose or orientation, this operator can be selected. For example, in automated assembly processes, ensuring components are aligned and installed at correct angles; checking if specific features on products are at correct rotation angles; classifying based on object orientation, etc.

Input Output

Input

Image: Single color image to be detected.

Output

Detection result: A detection instance list. Each element in the list represents an identified object, containing its class, confidence score and a rotated rectangle box describing its precise position and orientation.

Parameter Description

Weight File

Parameter Description

Load pre-trained model files for rotated object detection.

Parameter Adjustment

Select and upload model files trained for specific application scenarios and target objects. The model quality directly determines the accuracy and robustness of results.

Parameter Range

File format required is .pth format.

Confidence Threshold

Parameter Description

Used to set the "reliability" threshold for model output results. When the model identifies each object, it will give a score between 0 and 1, indicating how reliable the identification result is. Only objects with scores higher than this threshold will be output as valid results.

Parameter Adjustment

  • Increase value: Filtering conditions become stricter. The operator will only output results that the model is very confident about. This will reduce false identifications, but may miss some real targets with unclear identification features or partial occlusion.

  • Decrease value: Filtering conditions become more lenient. The operator will output more possible targets, but may introduce more false identifications. If many non-target objects are being incorrectly output, you can try increasing this threshold; if obvious targets are not being identified, you can appropriately decrease this threshold.

Parameter Range

[0,1], default value: 0.5

Enable GPU Acceleration

Parameter Description

Controls whether the operator uses CPU or GPU for computation. Since deep learning models have large computational requirements, using GPU can greatly improve processing speed.

Parameter Adjustment

  • Disabled (default): The operator will use CPU for computation. Although it has good compatibility and doesn’t require special hardware, processing one image may take longer time, not suitable for scenarios with high speed requirements.

  • Enabled (recommended): If the local device supports GPU graphics card, it’s recommended to enable this option, which will greatly improve processing speed.

Enable Rotation

Parameter Description

Sets whether the model is trained for angle rotation, used to rotate detection boxes to make them more closely fit targets.

Parameter Adjustment

  • Enabled: When the model is a rotated object detection model, rotation must be enabled. The operator can generate rotated detection boxes by parsing angle information output by the model.

  • Disabled: If the trained model is a standard object detection model that can only output horizontal rectangle boxes, this option needs to be disabled.

Legacy Annotation Model

Parameter Description

Used to handle models trained with legacy data annotation methods.

Parameter Adjustment

  • Enabled: If the model was trained using early "strict sequential four-point annotation" method, please enable this option.

  • Disabled: If the model was trained using standard "rotated rectangle" annotation method, please disable this option. If unsure whether it’s a legacy annotation model, you can try both settings. When detection box angles are obviously wrong (for example, angles are always offset by 90 degrees), you can switch this option to try solving the problem. Generally for newly trained models, keep this option disabled by default.