ICME26 Grand Challenge IDA^*: Cross-Scenario Defect Detection and Fine-Grained Severity Grading for High-Precision Manufacturing

The quality control of Printed Circuit Board Assemblies (PCBA) and high-density electronic components represents a cornerstone of modern manufacturing. As electronic devices become increasingly miniaturized, the automated inspection of solder joints, connector pins, and substrate surfaces faces unprecedented challenges. However, a critical bottleneck restricts the massive deployment of current Industrial AI: the lack of Cross-Scenario Generalization Ability and Fine-Grained Severity Grading Capability.

In real-world workflows, production lines and sensing environments frequently change. Conventional models often suffer from catastrophic performance degradation when transferred from one production line to another unseen domain. When transferred to another unseen scenario, existing benchmarks predominantly focus on binary decisions, lacking fine-grained assessment ability of the severity, which is crucial for better optimizing industrial pipelines.

IDA 2026 establishes two distinct competition tracks:

The goal is to advance AI systems from mere “defect finders” (Track 1) to intelligent “risk assessors” (Track 2) for electronics assembly.

• The training dataset will be sent to the contact email address provided in the registration form.

We define separate scoring metrics for each track. Crucially, both tracks require strong capabilities in Defect Localization and Classification, but with different weighting priorities.

Track 1: Cross-Scenario Defect Detection

The ranking for Track 1 is determined by the Detection Score \(S_{Track1}\). This score emphasizes the model's ability to find and recognize defects in new domains while controlling false alarms.

$$ S_{Track1} = 0.3 \times S_{loc} + 0.3 \times S_{cls} + 0.4 \times S_{screen} $$

• Localization (\(S_{loc}\)): Mean Intersection over Union (mIoU).
• Classification (\(S_{cls}\)): Macro-F1 Score for defect type identification.
• Screening Efficiency (\(S_{screen}\)): composite of Image-Level Recall and Specificity to penalize false alarms in unseen domains.

Track 2: Fine-Grained Severity Grading

The ranking for Track 2 is determined by the Risk Assessment Score \(S_{Track2}\). While severity grading is the primary goal, accurate localization and classification are prerequisites for valid reasoning.

$$ S_{Track2} = 0.2 \times S_{loc} + 0.2 \times S_{cls} + 0.6 \times S_{grade} $$

• Severity Grading (\(S_{grade}\)): Quadratic Weighted Kappa (QWK).
• Localization (\(S_{loc}\)) & Classification (\(S_{cls}\)): ensures the severity grade is assigned to the correct defect instance and type.

We provide official baseline codes to help participants get started quickly.

• Track 1 Baseline: Cross-Scenario Defect Detection Baseline
• Track 2 Baseline: Fine-Grained Severity Grading Baseline

The baseline is intended for reference only. Participants are encouraged to develop their own innovative solutions.

Participants can choose to compete in one or both tracks.

• Track 1 Submission: Download Example
• Track 2 Submission: Download Example

The submitted JSON file should follow the structure shown in the example files. The fields are defined as follows:

• results: total results of all test images.
• file_name: image name.
• width: width of the image.
• height: height of the image.
• defect_info: all defect information for one test image.
• category_id: category ID of one defect.
• severity: severity level of one defect.
• bbox: bounding box of one defect, formatted as [x_min, y_min, x_max, y_max]. All coordinates should be normalized to the range [0, 1].
• segmentation: segmentation points of one defect, formatted as [x1, y1, x2, y2, ..., xn, yn]. All coordinates should be normalized to the range [0, 1].

• Wei Sun, wsun@cee.ecnu.edu.cn
• Weixia Zhang, zwx8981@sjtu.edu.cn
• Linhan Cao, caolinhan@sjtu.edu.cn
• Xiongkuo Min, minxiongkuo@sjtu.edu.cn
• Xiaoping Zhang, xpzhang@ieee.org
• Patrick Le Callet, patrick.lecallet@univ-nantes.fr
• Guangtao Zhai, zhaiguangtao@sjtu.edu.cn