SortScrews: A Dataset and Baseline for Real-time Screw Classification

This paper introduces SortScrews, a specialized dataset and baseline framework for real-time screw classification in industrial automation, demonstrating that lightweight deep learning models can achieve strong accuracy even with a limited set of 560 images captured under controlled conditions.

The Gist

Imagine you are running a busy factory where a robot arm needs to sort a giant pile of mixed-up screws. To the human eye, a 2.5cm round-head screw looks very different from a 3.5cm flat-head screw. But to a computer camera, they can look almost identical—just a tiny, shiny metal cylinder. If the robot grabs the wrong one, the assembly line stops, or the machine breaks.

This paper is about teaching computers to be expert screw-sorters, but with a twist: they did it with very little data and very simple tools.

Here is the story of SortScrews, broken down into simple parts:

1. The Problem: The "Needle in a Haystack" of Data

Usually, to teach a computer to recognize things (like cats or dogs), you need millions of photos. It's like trying to learn a language by reading an entire library. But in a factory, you don't have millions of photos of every specific screw type. You might only have a few hundred.

Most existing datasets are like giant encyclopedias, but factories need a specific "pocket guide" for tiny, tricky parts. There was no good, free guide for screws, so the authors decided to make their own.

2. The Solution: A "Screw Photo Booth"

The authors built a simple, low-cost "photo booth" for screws.

• The Setup: They used a cheap webcam, a wooden stand, and a printed paper guide (like a target on the floor) to tell the screw exactly where to sit.
• The Process: They took 560 photos of 6 different types of screws.
• The Trick: They didn't just take perfect photos. They moved the light slightly and changed the camera angle a tiny bit. This is like taking a selfie in different lighting so you learn to recognize your face even when the sun is in your eyes.

They also wrote a free "recipe" (a script) so anyone else can build this photo booth with their own cheap camera and take pictures of their own weirdly shaped nuts and bolts.

3. The Test: Can a "Small Brain" Learn?

Usually, AI models are like giant supercomputers. But for a factory robot, you need something fast and lightweight, like a smartwatch instead of a mainframe.

The authors tested two "small brains" (AI models) on their screw photos:

1. ResNet-18: A classic, reliable, lightweight model.
2. EfficientNet-B0: A newer model designed to be super efficient.

The Result:

• ResNet-18 was the star player. It got 96.4% of the screws right. It was so fast it could sort about 155 screws per second.
• EfficientNet-B0 was a bit slower and got 86.2% right.

The Big Surprise: The "older," simpler model (ResNet-18) actually did a better job than the fancy new one. It proved that you don't need a massive supercomputer to sort screws if you control the environment well.

4. The "Oops" Moments (Failure Analysis)

Even the best student makes mistakes. The AI got confused when two screws looked too similar.

• The Mix-up: It sometimes confused a "Round-head 2.5cm" screw with a "Flat-head 3.5cm" screw.
• Why? To the camera, they are both just "metal cylinders." Without more angles (like seeing the screw from the side) or more photos, the AI couldn't tell the difference between the shape of the head and the length of the body.
• The Bias: The AI also got a little lazy. It started guessing based on where the screw was in the photo rather than what it actually looked like. It's like a student who memorizes the answer key's position on the page instead of learning the math.

5. Why This Matters

This paper is a gift to the robotics and factory world for three reasons:

1. The Dataset: They gave away 560 photos of screws for free, so other researchers can start training immediately.
2. The Blueprint: They showed that you don't need a million-dollar lab to build a training system; a webcam and a piece of wood work fine.
3. The Proof: They proved that with controlled lighting and simple setups, even small, fast AI models can do high-quality industrial work.

In a nutshell:
Think of this paper as a DIY guide for teaching a robot to sort screws. Instead of buying a expensive, complex system, they built a simple "screw photo booth," took a few hundred pictures, and showed that a modest AI model can learn to sort them almost perfectly. It's a reminder that sometimes, the simplest tools create the smartest solutions.

Technical Summary

Technical Summary: SortScrews - A Dataset and Baseline for Real-Time Screw Classification

1. Problem Statement

The paper addresses the critical need for automatic screw identification in industrial automation, robotics, and inventory management. While deep learning has advanced visual recognition, there is a significant scarcity of publicly available datasets for fine-grained industrial component recognition, specifically for screws.

Key challenges identified include:

• Subtle Geometric Variations: Screw types often differ only by minor details (head shape, length, thread pattern), making them difficult for computer vision systems to distinguish without controlled imaging.
• Data Scarcity: Industrial environments typically suffer from limited labeled data and constrained acquisition setups.
• Lack of Standardization: The absence of standardized datasets hinders the benchmarking of algorithms for automated sorting and robotic assembly.

2. Methodology

A. Dataset Construction (SortScrews)

The authors introduce SortScrews, a curated dataset designed for casewise visual classification under controlled conditions.

• Scale & Composition: The dataset contains 560 RGB images at 512 × 512 resolution.
• Classes: It covers 6 distinct screw categories (varying by head type and length) plus a background class to support rejection mechanisms. Each category is balanced with exactly 80 samples.
• Acquisition Setup:
  • Hardware: A low-cost setup using an iCAN C55N QHD 2K webcam, a wooden stand, and a printed guide for Point-of-View (POV) calibration.
  • Protocol: Images are captured with a single screw placed randomly within a calibrated region.
  • Variations: Data was collected under four different acquisition settings to introduce mild variations in lighting and camera perspective, simulating real-world industrial environmental changes while maintaining control.
• Reproducibility: A reusable data collection script is provided, allowing researchers to build similar datasets for custom hardware using inexpensive cameras.

B. Baseline Models & Training

To establish reference performance, the authors employed Transfer Learning using models pretrained on ImageNet.

• Architectures:
  • EfficientNet-B0: Selected for its parameter efficiency and scaling principles.
  • ResNet-18: A lightweight, widely used residual network.
• Training Configuration:
  • Optimizer: AdamW (Learning rate: $10^{-3}$, Weight decay: $10^{-4}$).
  • Input: Images resized to 224 × 224.
  • Hardware: Trained on a MacBook Pro (Apple M3) with Metal acceleration, demonstrating the feasibility of training on consumer-grade hardware.
  • Strategy: Backbone networks were optionally frozen to stabilize optimization given the small dataset size.

3. Key Contributions

1. SortScrews Dataset: A balanced, high-quality dataset of 560 images covering six screw types and background, specifically designed for fine-grained industrial classification.
2. Open-Source Pipeline: Release of a reusable, low-cost data collection script and hardware guide, enabling rapid dataset creation for custom industrial components.
3. Baseline Benchmarks: Comprehensive evaluation of EfficientNet-B0 and ResNet-18, establishing performance baselines and inference time metrics for real-time sorting applications.
4. Failure Analysis: Detailed investigation into model errors, highlighting specific confusion patterns between visually similar screw types.

4. Experimental Results

Classification Performance

Despite the small dataset size, the models achieved strong accuracy, validating the efficacy of controlled acquisition conditions:

• ResNet-18: Achieved 96.4% validation accuracy.
• EfficientNet-B0: Achieved 86.2% validation accuracy.
• Observation: Contrary to expectations on large-scale tasks, the lighter ResNet-18 outperformed EfficientNet-B0 in this specific fine-grained context.

Inference Speed (Real-Time Capability)

The models demonstrated suitability for real-time sorting systems:

• ResNet-18: 6.42 ms average inference time (155.8 fps on NVIDIA GPU).
• EfficientNet-B0: 17.95 ms average inference time (55.7 fps on NVIDIA GPU).

Failure Analysis

• Confusion Patterns: Errors were primarily concentrated between visually similar categories (e.g., confusing Round-head 2.5 cm with Flat-head 3.5 cm).
• Bias: The models exhibited an unexpected bias toward the spatial location of the screw in the frame, likely due to sparse semantic supervision. The authors suggest that adding bounding box supervision could mitigate this.

5. Significance and Future Work

• Industrial Impact: The paper demonstrates that low-cost, controlled acquisition setups can yield high-accuracy classification results even with small datasets, making automated sorting accessible to smaller manufacturers.
• Research Enabler: By releasing the dataset and collection tools, the authors lower the barrier to entry for research in industrial object recognition.
• Future Directions: The authors propose extending the dataset with multi-view capture, conveyor-belt environments, and depth/3D information to further support robotic manipulation and complex sorting tasks.

In conclusion, SortScrews fills a critical gap in industrial vision datasets, proving that standardized, controlled data collection combined with transfer learning is a viable and effective strategy for solving fine-grained mechanical component classification problems.