PaperlayerView exact revision on arXiv

Source-linked AI summary

MCFR: A Mask-Guided Coarse-to-Fine Regression Framework for Robust Multi-Variant Board-to-Board Connector Assembly

Guanghui Shen, Song Wang, Dan Wu

arXiv:2610.09327v1cs.RO

TL;DR

Multi-variant BTB connectors are difficult to localize robustly because their morphology and appearance vary, while fixed-view training and deployment can induce background spurious correlation. MCFR combines an object-aware mask prior with coarse-to-fine regression and photometric refinement. Across the reported evaluations, it outperforms representative baselines and achieves 99.25% average real-world insertion success on eight connector variants.

  • Problem

    Multi-variant BTB connectors pose a generalization challenge, and fixed-view training–deployment mismatch can make pose models rely on background displacement cues rather than object features.

    Source: §I · p. 1

  • Method

    MCFR combines an object-aware probability mask with coarse pose regression and explicit photometric refinement to suppress background interference and refine alignment.

    Source: §I · p. 1, 1) Object-Aware Mask Generation (OAMG) · p. 3, §V · p. 7

  • Results

    99.25% average insertion success rate over 8 connector variants, while MCFR outperforms representative baselines in translation and rotation accuracy.

    Source: §I · p. 2, §V · p. 7

  • Takeaways & Limitations

    Batch BTB insertion results and a real smartphone assembly task demonstrate the pipeline’s practical feasibility for multi-variant connector assembly.

    Source: §V · pp. 7–8

Abstract

from arXiv · show

Automated insertion of board-to-board (BTB) connectors in 3C manufacturing requires both high visual accuracy and strong deployment robustness. This problem remains challenging because multi-variant connectors exhibit significant morphological and appearance variations, making stable cross-variant generalization difficult, while the mismatch between training and deployment under fixed-view inspection settings induces background spurious correlation and degrades real-world performance. To address these issues, this paper proposes MCFR, a Mask-Guided Coarse-to-Fine Regression framework for multi-variant BTB connector assembly. By introducing an object-aware mask prior and explicit photometric refinement, the proposed method suppresses background interference and improves alignment accuracy and robustness in practical deployment. Experiments on a self-constructed multi-variant dataset, a BTB batch insertion testbed, and a real smartphone assembly task show that MCFR consistently outperforms representative baselines and achieves an average real-world insertion success rate of 99.25%. These results demonstrate the effectiveness and practical potential of MCFR for automated assembly of multi-variant BTB connectors.

I. INTRODUCTION

Multi-variant BTB connectors demand precise, robust pose detection, but appearance differences and fixed-view training–deployment mismatch hinder generalization. MCFR combines object-aware masks and coarse-to-fine regression with photometric refinement; experiments report improved baseline performance and 99.25% average real-world insertion success.

I. INTRODUCTION

  • BTB connector insertion requires pose accuracy within stringent tolerances, while variant diversity, miniature size, appearance changes, and weak localization features make reliable detection difficult.The stated typical tolerances are ±0.2 mm and ±1.5 °.Source: §I · p. 1
  • Fixed-view training can couple target and background movement, whereas deployment errors arise mainly from grasping and the background remains static, creating background spurious correlation.The paper describes this as a task-specific manifestation of shortcut learning under distribution shift.Source: §I · p. 1
  • MCFR uses a SAM-generated object-aware probability mask with a ViT-based coarse pose regressor, then applies robust pixel-level photometric refinement.The mask indicates connector localization regions to suppress background interference, while refinement aligns the coarse pose.Source: §I · pp. 1–2
  • 99.25% average real-world insertion success rate, with MCFR outperforming existing baselines in pose error and insertion success.Evaluations span a self-constructed multi-variant dataset, a batch insertion test platform, and a real smartphone assembly scenario.Source: §I · p. 2

II. RELATED WORK

Related work divides vision-based pose detection for flexible assembly into feature-engineered analytical methods and end-to-end regression. Analytical methods have advanced feature extraction with deep detectors, but their object-level representations and feature-design demands leave fine geometry and cross-variant generalization challenging.

A. Analytical Pose Estimation Based on Feature Engineering

  • Analytical pose estimation extracts image features and combines them with geometric or statistical analysis to calculate relative pose.Examples include PCA, LDA, and geometric solvers; traditional operators such as Canny and Hough depend heavily on empirical thresholds.Source: A. Analytical Pose Estimation Based on… · p. 2
  • Deep detection models such as YOLO improve localization-feature acquisition, but their object-level representations have limited capacity to describe fine geometric details.Traditional image-processing operators also have limited robustness to complex backgrounds and multiple part categories.Source: A. Analytical Pose Estimation Based on… · p. 2
  • Feature-engineering pipelines are constrained by handcrafted-feature design and extraction quality, making precision and generalization difficult to balance across diverse BTB connectors.Examples include edge-based alignment methods with strict edge-quality requirements and connector-specific pin-layout methods.Source: A. Analytical Pose Estimation Based on… · p. 2

B. End-to-End Pose Regression and Perception Enhancement

End-to-end pose regression reduces manual feature design and represents complex appearances, but category- and sensor-specific models can degrade on out-of-distribution samples. For fixed-view BTB assembly, the paper proposes combining object-aware mask priors, coarse regression, and explicit optimization; its problem formulation estimates relative pose between inspection images and converts it into robot compensation.

B. End-to-End Pose Regression and Perception Enhancement

  • End-to-end methods map sensory inputs directly to pose deviations or control commands, avoiding explicit feature design and modeling complex object appearances.Prior work includes Siamese CNN regression, digital-twin compensation for occlusion, and visuotactile approaches.Source: B. End-to-End Pose Regression and… · p. 2
  • Many end-to-end models are tailored to specific object categories and sensors, lack explicit spatial constraints, and can degrade under out-of-distribution samples or changing backgrounds and objects.The paper proposes a three-stage combination of object-aware mask priors, coarse regression, and explicit optimization for fixed-view multi-variant BTB detection.Source: B. End-to-End Pose Regression and… · pp. 2–3

A. Problem Statement

  • Relative pose is the object’s pose difference between two inspection images; the formulation avoids accumulating visual and robot absolute-position errors during insertion.Robot repeatability is about ±0.03 mm, compared with empirical absolute positioning error around ±0.1 mm and typical mating tolerance of ±0.2 mm.Source: A. Problem Statement · p. 3
  • The estimated in-plane translations and optical-axis rotation are converted through a calibrated coordinate transformation into an insertion compensation command for the robot end-effector.The pipeline records a nominal-grasp reference image and estimates the current image’s planar pose relative to it.Source: A. Problem Statement · p. 3

B. MCFR: Mask-Guided Coarse-to-Fine Regression

MCFR estimates relative pose through object-aware mask generation, mask-guided coarse regression, and photometric refinement. Its mask is built from filtered SAM responses to emphasize connector localization cues and suppress background interference.

B. MCFR: Mask-Guided Coarse-to-Fine Regression

  • MCFR estimates relative pose in three stages: object-aware mask generation, cross-attention coarse regression, and robust photometric refinement.The stages construct a visual prior, produce an initial pose, and refine it through image-domain alignment.Source: B. MCFR: Mask-Guided Coarse-to-Fine… · p. 3, Figure 2 · p. 4

1) Object-Aware Mask Generation (OAMG):

  • SAM responses to uniformly distributed grid prompts provide target probability maps without task-specific fine-tuning.The maps are used to construct the object-aware visual prior.Source: 1) Object-Aware Mask Generation (OAMG) · p. 3
  • Retained maps are filtered for rectangularity and connector-consistent spatial extent, then summed pixelwise and tanh-normalized to highlight the connector and suppress background interference.The filtering reflects the approximately rectangular distribution of task-relevant localization cues in the fixed-view setting.Source: 1) Object-Aware Mask Generation (OAMG) · p. 3

2) Cross-Attention Coarse Regression (CACR):

The coarse regression stage masks the reference and current images before feature encoding. Mutual cross-attention models their correspondence, and an MLP predicts the initial relative pose.

2) Cross-Attention Coarse Regression (CACR):

  • Pixel-wise multiplication applies the object-aware masks to the reference and current images, enhancing target responses while suppressing background.The masked images are inputs to the feature encoder.Source: 2) Cross-Attention Coarse Regression… · p. 4
  • A ViT-based backbone encodes the masked images, with MAE pre-training used to improve training efficiency and feature quality.The paper motivates ViT by its structural scalability and suitability for connector appearance variations.Source: 2) Cross-Attention Coarse Regression… · p. 4
  • Mutual cross-attention models correspondence between reference and current features, and fused features feed an MLP regression head to estimate relative pose.This coarse estimate provides the initialization for subsequent refinement.Source: 2) Cross-Attention Coarse Regression… · p. 4

3) Robust Photometric Refinement (RPR):

Robust photometric refinement optimizes the coarse pose by aligning a warped current image with the reference. The refined relative pose is used for pose compensation before insertion.

3) Robust Photometric Refinement (RPR):

  • The coarse estimate initializes optimization, which applies a planar rigid-body transformation to the current image in the image domain.A more accurate relative pose should yield higher photometric consistency between the warped current image and reference image.Source: 3) Robust Photometric Refinement (RPR) · p. 4
  • NCC-based similarity between the warped and reference images is used for alignment and is more robust than direct pixel-wise error to global illumination shifts and local contrast changes.The refinement is introduced to address deployment effects including illumination variation, background disturbance, and cross-category distribution shift.Source: 3) Robust Photometric Refinement (RPR) · pp. 4–5
  • The refined relative pose is used for pose compensation before insertion, combining deep regression with explicit optimization to improve detection stability and alignment precision in real assembly scenarios.The paper states this combination preserves cross-category adaptability.Source: 3) Robust Photometric Refinement (RPR) · p. 5

A. Overview

The experiments cover dataset evaluation, batch insertion on a BTB mating testbed, and a smartphone camera-module assembly task. The setup combines robot grasping and visual localization with high-precision relative-pose detection.

A. Overview

  • Evaluations cover a self-constructed multi-variant dataset, BTB batch insertion, and a practical smartphone camera-module insertion task.The smartphone task is used to verify feasibility in a complete assembly pipeline.Source: A. Overview · p. 5

1) Hardware:

  • The hardware includes a UR5e robot, EPick vacuum gripper, grasp-stage and detection cameras, and a 10-socket BTB mating testbed.The detection camera is a 3072 × 2048 monocular camera; a Redmi K40 and camera modules support the realistic assembly test.Source: 1) Hardware · p. 5, Figure 3 · p. 5

2) Dataset Collection:

  • The dataset contains 4,500 samples across 10 BTB connector variants and 3 smartphone camera modules, split 8:1:1 into training, validation, and test sets.A simulated deployment test set composites segmented connectors onto a separately captured fixed-station background so only connector pose varies.Source: 2) Dataset Collection · p. 5

3) Model Configuration and Training:

MCFR achieves the lowest mean translation and rotation errors on validation and test sets, and ablations show gains from both its object-aware mask and photometric refinement.

3) Model Configuration and Training:

C. Algorithm Comparison and Ablation Study

  • Evaluation uses Trans. MAE, Rot. MAE, and maximum errors, comparing YOLO+LDA with ResNet50+LSTM and MCFR.YOLO+LDA combines YOLO pin segmentation with LDA-based geometric analysis; ResNet50+LSTM is an end-to-end Siamese regression model.Source: C. Algorithm Comparison and Ablation… · p. 5
  • 0.027 mm Trans. MAE and 0.126° Rot. MAE make MCFR best on test, versus ResNet50+LSTM at 1.062 mm and 1.335°.MCFR also records validation Trans. MAE of 0.013 mm and Rot. MAE of 0.015°, with maximum translation errors of 0.064 mm on validation and 0.196 mm on test.Source: Table I · p. 6, Figure 4 · p. 6, C. Algorithm Comparison and Ablation… · p. 6
  • Removing the object-aware mask raises test Trans. MAE from 0.027 mm to 0.793 mm, while removing refinement raises test Trans. MAE to 0.052 mm and Rot. MAE to 0.257°.The mask ablation is associated with sensitivity to fixed-background deployment; the full model relies more on connector-intrinsic cues.Source: C. Algorithm Comparison and Ablation… · p. 6
  • With the mask prior, occlusion sensitivity concentrates on the connector body and pin region, whereas the unmasked model attends more to image boundaries and the gripper contour.Fig. 5 contrasts the model without the mask in the top row and with the mask prior in the bottom row.Source: C. Algorithm Comparison and Ablation… · p. 6, Figure 5 · p. 6

D. Real-World Insertion Experiments

Across eight BTB connector variants, MCFR has the highest average insertion success rate, and it also completes the smartphone camera-module insertion pipeline under tighter structural constraints.

D. Real-World Insertion Experiments

  • Each of eight connector variants receives 50 full trials, with success counted only when assembly finishes before insertion-direction contact force reaches 25 N.The threshold is based on the average peak insertion force observed during successful manual alignment and insertion.Source: D. Real-World Insertion Experiments · p. 6
  • 99.25% average insertion success rate makes MCFR best across eight variants, versus 94% for MCFR w/o RPR and 22% for ResNet50+LSTM.ResNet50+LSTM performs worse across variants, with a more noticeable drop on connectors with larger pin counts.Source: Figure 7 · p. 7, D. Real-World Insertion Experiments · p. 7
  • Connector labels encode pin pitch in their first two digits and pin count with the suffix P; higher pin counts are more prone to jamming under slight alignment errors.For example, 35 denotes 0.35 mm pitch and 24P denotes 24 pins.Source: D. Real-World Insertion Experiments · p. 7
  • The deployment mismatch is fixed background versus training images collected with randomized part motion, which can cue end-to-end models to background displacement.MCFR suppresses this spurious correlation with its mask prior and uses photometric refinement for alignment precision.Source: D. Real-World Insertion Experiments · p. 7
  • The smartphone task successfully executes recognition, grasping, detection, alignment, insertion, and homing despite tighter structural constraints and higher assembly difficulty.This task evaluates transferability in a real smartphone camera-module insertion scenario.Source: D. Real-World Insertion Experiments · p. 7, Figure 6 · p. 7

V. CONCLUSIONS

MCFR combines an object-aware mask prior, mutual cross-attention coarse regression, and explicit photometric refinement to improve alignment accuracy and deployment robustness. It reports strong dataset and real-world insertion results, while zero-shot initialization for entirely unseen connector variants remains unaddressed.

V. CONCLUSIONS

  • An object-aware mask prior, mutual cross-attention coarse regression, and explicit photometric refinement suppress background spurious correlation while improving alignment accuracy and deployment robustness.Source: §V · p. 7
  • 99.25% average insertion success over 8 connector variants was achieved on the BTB mating testbed.A real smartphone assembly task further demonstrates the complete pipeline's feasibility in practical deployment.Source: §V · pp. 7–8
  • MCFR outperforms representative analytical and end-to-end baselines in translation and rotation accuracy on the self-constructed dataset.Ablation studies and occlusion sensitivity visualization also validate the proposed design from quantitative and feature-attribution perspectives.Source: §V · p. 7
  • MCFR is presented as extendable to other object categories without architectural modification, as a practical visual solution for high-precision robotic assembly.Source: §V · p. 8
  • Zero-shot initialization on entirely unseen connector variants is not explicitly addressed and remains future work.Source: §V · p. 8
Loading 2610.09327v1…