Vision Parameter Configuration
About 2050 wordsAbout 7 min
5. Vision Parameter Configuration
This chapter verifies parameters in the order they are processed: first confirm instance recognition and point clouds, then check sorting, keypoint coarse matching, and point cloud fine matching, and finally verify pick point filtering and output.
Adjust only one parameter or one group of parameters related to the current issue at a time. Click Apply and retest with the same historical data. Do not change the model, template, and multiple groups of filtering thresholds at the same time, or it will be difficult to determine what caused an improvement.
5.1 Establish an Initial Validation Configuration
- Select the target, camera, ROI, end-effector, and calibration results configured in the previous two chapters.
- Open the vision parameter configuration generated when the task was created.
- First run a representative sample of clearly visible, unobstructed incoming parts with the default parameters. Record the number of recognized instances, pose results, and processing time.
- Use the following checklist to identify which stage needs adjustment.
| Stage | What to check | Condition for proceeding to the next stage |
|---|---|---|
| 2D instance recognition | Whether each cylinder is recognized separately and whether targets at the edges are missed | Targets are complete, and adjacent cylinders have not been merged |
| Instance point clouds and sorting | Whether each point cloud belongs to the same target and whether the order meets process requirements | Valid features are retained, and the sorting reference is clear |
| Pose estimation | Whether the template fits and the orientation is correct | Coarse matching is correct, and fine matching meets the requirements |
| Pick point processing | Position, orientation, and reasons for filtering | Normal results are retained, and abnormal results are removed |
5.2 Adjust 2D Recognition
5.2.1 Preprocessing and Instance Segmentation
First check that the ROI, original image, and model are correct, then adjust the recognition parameters.
- Scale factor: Start with the default value. If a cylinder is missed, one box contains multiple cylinders, or a box is incomplete, adjust the value gradually while observing samples in both central and edge positions.
- Minimum confidence threshold: Lowering it can retain more low-confidence results but may increase false recognition. Raising it can exclude some incorrect results but may also miss real cylinders. An increase in the number of instances alone is not enough to judge the result.
- Edge enhancement: Evaluate this only when the boundaries between adjacent cylinders are unclear and the depth data is reliable. Check whether processing mistakenly removes cylinder surfaces or other valid features.
- Remove image background outside ROI 3D: Evaluate this when environmental interference is significant. After enabling it, check whether target contours remain complete and whether depth holes create abnormal background regions.
For general parameter descriptions, see the 2D Instance Segmentation Configuration Guide. For preprocessing methods, see the 2D Preprocessing Function Configuration Guide.
5.2.2 Enable Auto Augmentation as Needed
Enable auto augmentation only when a single scale factor cannot cover all normal variations in incoming parts:
- Select scale factors that have been verified as effective and the rotation angles that need to be covered.
- For general target scenarios, select the axis-aligned bounding box type AABB as required by the corresponding interface.
- Set the original scale factor to
1.00, and configure the scale factors to be combined in the auto augmentation items. - Add combinations one at a time. Compare recognition completeness, false recognition, and processing time, and remove combinations that provide no consistent benefit.
More augmentation combinations are not necessarily better. If the actual cylinders do not match the model configuration, correct the model or training data first instead of adding inference combinations indefinitely.
5.3 Generate Instance Point Clouds and Set the Order
5.3.1 Select an Instance Point Cloud Generation Method
- If the mask is accurate and covers the matching region, evaluate the mask method first to reduce the inclusion of background point clouds and point clouds from adjacent cylinders.
- If the mask cuts off necessary geometric features, evaluate the bounding box method, but check whether it includes adjacent cylinders or the pallet.
- Use “After segmentation” when instances do not need to be filtered in advance. Evaluate “After filtering” when invalid instances must be excluded first to reduce subsequent computation.
Check the instance point clouds actually generated each time you switch methods; do not rely only on the 2D bounding boxes. When using the mask method, retain the mask output it depends on. For specific steps, see the Point Cloud Generation Configuration Guide.
5.3.2 Set Instance Filtering and Sorting
- Set the necessary instance filters according to the target’s actual dimensions, valid region, or recognition quality. First confirm that normal cylinders are not filtered out.
- If a picking order is required, set the reference coordinate system for sorting first.
- Configure the X/Y strategies according to the actual row and column directions. If incoming parts have multiple layers, also group them by height and specify which layer should be processed first.
- Use samples from corners and edges, with missing parts and variations in height between layers, to check whether instance IDs follow the expected order.
- If the process does not allow a failed position to be skipped in favor of the next position, also check Pose Calculation Exception Alarm in the vision calculation configuration.
Do Not Guess the Sorting Direction from the Image
The positive Z direction of the ROI coordinate system may differ from that of the camera coordinate system. “Z from small to large” does not necessarily mean “top to bottom.” Confirm the coordinate axis directions first, then check the actual sorting results.
For configuration methods, see the Instance Filtering Function Configuration Guide, Instance Sorting Configuration Guide, and Vision Calculation Configuration Function User Guide.
5.4 Verify General Pose Estimation
5.4.1 Check 3D Preprocessing
First confirm that the instance point cloud within the ROI belongs to one cylinder. When using distance filtering, denoising, or downsampling, pay particular attention to whether directional features and contours are still retained.
If preprocessing leaves only a plane without directional features, a score from subsequent matching does not establish that the orientation is correct. For parameter descriptions, see the 3D Preprocessing Function Configuration Guide.
5.4.2 Verify Keypoint Coarse Matching First
- Check that the target’s keypoint file, visual model, and normalized mesh correspond to one another.
- Check whether the initial pose falls on the correct cylinder and whether its orientation matches the physical target.
- If coarse matching fails, first check instance segmentation, keypoint configuration, and point cloud completeness before considering an adjustment to the Coarse Matching Evaluation Threshold.
- If coarse matching is clearly misaligned or the front and back are reversed, do not treat it as a correct result merely by relaxing the threshold.
In the V1.9.0 General Pose Estimation Guide, the default coarse matching evaluation threshold is 10 mm. A larger threshold permits greater keypoint error and may also accept incorrect matches.
5.4.3 Then Verify Point Cloud Fine Matching
Once the position and orientation from coarse matching are generally correct, enable Object Pose Correction and check how well the template fits the instance point cloud.
- Fine Matching Search Radius: The default value in the V1.9.0 guide is 10 mm. A radius that is too small may fail to find enough point pairs; one that is too large may introduce incorrect corresponding points.
- Fine Matching Search Mode: Compare point-to-point, point-to-plane, or a combination of both. The combined mode increases computation, so compare both error and cycle time.
- Contour Mode: Evaluate this when contour features are stable and matching needs improvement. If the contour itself is missing or contains noise, correct the input first.
- Save Fine Matching Data: Enable this during diagnosis to compare results before and after correction. For production operation, set it according to traceability and storage requirements.
Parameter Units
Older materials may use meters, while the distance parameters above use millimeters in the current guide. Check the units shown in the interface before entering values; do not copy old version values directly into new version fields. Do not assume that all other parameters use the same units.
For the complete parameters and limitations, see the General Pose Estimation Configuration Guide.
5.5 Configure Pick Point Processing
First confirm that the target’s pick points are correct, then set the filtering conditions for runtime results.
| Check objective | Corresponding processing | Validation method |
|---|---|---|
| Exclude low-quality poses | Filter based on fine matching scores | Compare both correct and incorrect match samples; do not focus only on the number of outputs |
| Limit the approach direction | Filter based on the picking pose angle range | Check the ROI axes, positive and negative directions, and picking pose axes |
| Limit output positions | Filter pick points outside the ROI 3D type region | Use samples inside and outside the boundary to validate the workspace or picking region |
| Check risks around the gripper | Filter pick points where the target collides with the gripper | Check the gripper’s dimensions, coordinates, and mounting pose, and inspect collision-free and borderline samples |
| Incoming parts with multiple layers or occlusion | Filter pick points on occluded targets | Verify that results for targets below are excluded while the upper layer has not been removed |
Models used for collision detection should meet the simplification requirements of the corresponding function while retaining the actual gripper’s key geometry and safety margin. Do not shrink the model or remove obstacles to make a collision check pass, and do not publish site-specific attachments directly as general examples.
For parameter and model requirements, see the Pick Point Filtering Function Configuration Guide. If pick points must be adjusted or reordered during operation, see the Pick Point Adjustment Function Configuration Guide and Pick Point Sorting Configuration Guide, respectively.
5.6 Apply and Retest
- Click Apply and confirm that the task has selected the vision parameters modified in this session.
- Run fixed historical samples and compare the results and processing time at each stage.
- If instances are present but there is no output, check whether pose estimation failed or the pick points were filtered out. Do not simply disable filtering.
- Proceed to operation and acceptance after normal samples, edge samples, samples with missing parts, and samples with permitted pose variations all pass.
Return to Case Overview · Previous Chapter: Point Cloud Template and Target Configuration · Next Chapter: Operation and Acceptance