cvtechniques/Motherboard-Parts-Locator

Motherboard Part Locator

blah blah blah. not finished.

Training Data

Forked From: GradResearch Computer Vision Model - https://universe.roboflow.com/gradresearch/gradresearch

Classes: 5 Images: 95

Dataset: Motherboard Part Locator (v5) - https://universe.roboflow.com/bis497/motherboard-part-locater

Classes: 4 Images (preaugmentation): 206 Images (post): 472

Class Name
CPU Socket
PCIe Slot
Power Connector
RAM Slot

Final Class Distribution

Class	Annotations
CPU Socket	205
PCIe Slot	444
Power Connector	320
RAM Slot	221

Data Collection Methodology

1. Image Collection: An R script was created to scrap 500 bare motherboard images from Bing images. This was created using 'httr', 'rvest', 'magick', and 'digest'. This was then manually reviewed and reduced to 101.
2. Locating More Images: The dataset 'GradResearch Computer Vision Model' was found on Roboflow. This would be the forked dataset that would be used as a base.
3. Label Standardization: The labels of the forked dataset were changed and standardized. An extra 'cpu' was removed and correctly labeled, leaving 4 classes.
4. Annotation: Half of the new 101 images were manually annotated. The later half was annotated with the forked model, then later reviewed again. To note, the forked model could not identify CPU Sockets.
5. Preprocessing: Preprocessing was applied; all images are now 640x640.

Note: The dataset contained mixed detect/segment annotations inherited from the original fork. Only bounding box annotations were used during final training.

Dataset Splits

Split	Images (pre-augmentation)	Split %
Train	134 → 399 augmented	65%
Validation	49	24%
Test	24	12%

Augmentations (applied to expand dataset to 472 images):

• Horizontal flip
• Rotation ±15°
• Blur up to 1.3×
• Noise

Training Procedure

Hardware: A100 Training time: ~0.07 hours
Platform: Google Colab
Epochs: 100 Patience: 50

Evaluation Results

Overall Metrics

Metric	Score
Precision	0.889
Recall	0.761
mAP@50	0.827
mAP@50-95	0.559

Per-Class Metrics (Validation)

Class	Precision	Recall	mAP@50	mAP@50-95
CPU Socket	0.967	0.894	0.963	0.678
PCIe Slot	0.878	0.602	0.690	0.417
Power Connector	0.811	0.755	0.815	0.560
RAM Slot	0.899	0.792	0.839	0.592

Comparison: Original (Forked) vs. This Model (v5)

Metric	Original (Forked)	v5 (This Model)
Images	195 (150/30/15)	472 augmented
Split	77/15/8%	65/24/12%
mAP@50	91.9%	82.7%
mAP@50-95	N/A	55.9%
Precision	87.7%	88.9%
Recall	88.2%	76.1%

The lower mAP@50 compared to the original fork is expected. The success criteria set (mAP@50 > 75% and Precision > 60%) were met!

Limitations and Biases

• PCIe Slot performance is the weakest class (mAP@50 = 0.69, Recall = 0.60). The confusion matrix shows PCIe Slots are frequently missed (35 background false negatives) and occasionally confused with Power Connectors. This is likely because the scarcity of close-up or angled PCIe Slot images and itr's confusion with the RAM.
• Small dataset: ~206 raw images is a limited training set.
• Only 4 classes: A beginner builder also needs to identify storage connectors (M.2, SATA), fan headers, front panel connectors, and I/O ports. This model does not detect those and will need them to be added to be more useful to a beginner PC builder.
• Nano-sized for speed: The YOLOv11n trades accuracy for speed.