💻 computer science

Counting Machine Parts

This paper presents an enhanced FamNet approach with an additional loss component for counting machine washer parts, which outperforms traditional image processing, instance segmentation, and density map estimation baselines by achieving a Mean Absolute Error of 1.96.

Original authors: Benedict Florance Arockiaraj, Elizabeth Dinella, Ankit Billa, Ajay Anand

Published 2026-05-19

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Benedict Florance Arockiaraj, Elizabeth Dinella, Ankit Billa, Ajay Anand

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are standing in front of a giant, tangled pile of metal washers. They are stacked on top of each other, some are hidden behind others, and the lighting is a bit dim. Your job is to tell someone exactly how many washers are in that pile.

This paper is about building a "smart camera" that can do this counting job automatically. The researchers tried several different ways to solve this puzzle, eventually finding a method that works very well.

Here is how they did it, explained simply:

The Problem: Why is this hard?

If you try to count these washers using old-school computer tricks (like looking for edges or circles), it's like trying to count people in a crowded concert by only looking at their shoes. The washers overlap, they are different sizes, and shadows make it confusing. The computer gets lost and counts the same washer twice or misses one entirely.

The Solution: A "Density Map" Approach

Instead of trying to find and count every single washer one by one (which is like trying to pick out every specific person in that crowd), the researchers taught the computer to look at the whole pile and guess the "density" of the washers.

Think of it like looking at a pile of sand. You don't count every grain; you look at how high the pile is and how wide it is to estimate the total amount. The computer creates a "heat map" of the image: bright spots mean "lots of washers here," and dark spots mean "no washers here." By adding up all the brightness, it gets the total count.

The Tools They Used

The team tested three different "tools" to solve the problem:

The Old School Tool (Image Processing): This is like using a ruler and a magnifying glass. The computer tries to find perfect circles. It worked okay, but because the washers were piled up and messy, the computer got confused and made many mistakes (counting about 29 washers off on average).
The "Spotter" Tool (Mask-RCNN): This is a standard AI that tries to draw a box around every single washer it sees. It's like a security guard trying to tag every person in the crowd. It got better when they trained it specifically on washers, but it still struggled when washers were hiding behind each other.
The "Heat Map" Tool (FamNet): This is the star of the show. It's based on a system called FamNet. Instead of boxing items, it learns to paint a density map. The researchers tweaked this system to make it even better for their specific pile of washers.

The Secret Sauce: Two Improvements

The researchers made two clever adjustments to the "Heat Map" tool to make it a champion:

The "No-Go Zone" Rule (Mismatch Loss): Sometimes, the computer gets too excited and paints a "hot spot" (predicting a washer) where there is actually just empty black background. The researchers added a special rule: "If you see a spot where there is definitely no washer, you get a penalty." This stopped the computer from hallucinating extra washers.
The "Group Vote" (Angle Aggregation): The dataset was special because every pile of washers was photographed from 9 different angles. Imagine asking 9 different people to count the same pile from different sides. The researchers found that if they took the average of all 9 guesses, the result was much more accurate than any single guess. It's like taking a class vote; the average is usually closer to the truth than one person's guess.

The Results

When they put everything together, their final system was incredibly accurate.

The old methods were off by about 29 washers on average.
The "Spotter" tool was off by about 4 washers.
Their final "Heat Map" system was only off by 1.96 washers on average.

What Didn't Work?

They also tried to use "depth" technology (like 3D glasses) to separate the washers into layers (front, middle, back) so they could count them easier. However, the lighting and the way the washers were piled made this 3D separation messy. It ended up causing the computer to double-count things, so they abandoned this idea.

The Bottom Line

The paper concludes that while this specific AI is great at counting these specific metal washers, it's a very specialized tool. It's like a master chef who can make the perfect omelet but might not know how to bake a cake. To make it a "master chef" for any object (cars, trees, people), you would need to train it on a much wider variety of things. But for this specific job of counting tangled metal parts, they found a solution that works almost perfectly.

Technical Summary: Counting Machine Parts

Problem Definition

The paper addresses the challenge of counting densely packed objects within images, specifically focusing on industrial machine washer parts. While object detection systems have advanced, the authors note that state-of-the-art detection models struggle with high-density scenes due to overlapping objects, occlusions, varying scales, and lighting conditions. The specific dataset consists of 8,100 images of washer bundles captured from nine different angles, with an average of approximately 26 objects per image. The primary difficulty lies in distinguishing individual circular parts when they are heavily occluded or clustered, a scenario where traditional counting methods and standard detection architectures often fail.

Methodology

The authors propose a deep learning-based approach that extends the FamNet architecture [12], a few-shot regression model designed to learn a density map for object counting. The methodology involves several key components:

1. Network Architecture and Adaptation

The core model utilizes a ResNet-50 backbone pre-trained on ImageNet for multi-scale feature extraction. Unlike the original FamNet, which uses a few exemplar bounding boxes to adapt to different classes at test time, this implementation is tailored for a single class (washer parts). Consequently, the model is provided with all bounding boxes for the washers in an image rather than a few exemplars. These boxes undergo ROI pooling on the convolutional feature maps to generate multi-scale features, which are then fed into a density prediction module consisting of five convolution blocks and three upsampling layers to produce a 2D density map. The final object count is derived by summing the values of this density map.

2. Ground Truth Density Construction

To generate ground truth density maps, the authors deviate from point-wise annotations used in previous works. Instead, they anchor point annotations to the center-of-mass of the object polygons. A dynamic window size for Gaussian smoothing is calculated based on the average distance between each point and its 1-nearest neighbor, ensuring the density map adapts to the specific spatial distribution of the washer parts.

3. Mismatch Loss

A novel contribution is the introduction of a Mismatch Loss function. The standard Mean Squared Error (MSE) loss in FamNet only penalizes the total error between predicted and ground truth density maps. The proposed Mismatch Loss specifically penalizes pixels in the predicted density map that correspond to areas in the image where no object exists (based on binary masks derived from annotations). Formally, this loss sums the predicted density values where the ground truth mask indicates the absence of an object. This loss is combined with the standard density MSE loss using a weighting hyperparameter $\lambda = 10^{-9}$ .

4. Angle Aggregation

Since the dataset includes nine different viewing angles for each scene, the authors employ a post-processing step to aggregate predictions. They experiment with max, min, and average aggregation methods across the nine angles. Empirical results indicated that taking the average of the predicted counts yielded the best performance.

5. Alternative Approaches (Depth Separation)

The paper also explores depth-based separation to handle occlusions, attempting to split images into depth layers using Multi-Level Otsu Thresholding and Dense Prediction Transformers (DPT). However, these experiments were largely unsuccessful due to the inability to cleanly separate overlapping objects based on grayscale intensity or depth maps, leading to duplicate counts and increased error.

Key Contributions

Adaptation of FamNet: Successfully adapted a few-shot counting model to a single-class, high-density industrial counting task by modifying the input mechanism to use full bounding box sets.
Mismatch Loss: Proposed an additional loss function designed to penalize false positives in the density map (predicting objects where none exist), though its efficacy was limited in this specific dataset context.
Multi-Angle Aggregation: Demonstrated that averaging predictions across multiple viewing angles of the same object bundle significantly reduces counting error compared to single-view predictions.
Comprehensive Baseline Comparison: Evaluated the proposed method against traditional image processing pipelines (Hough Transform) and instance segmentation (Mask-RCNN), highlighting the limitations of these approaches in dense scenarios.

Results

The performance of the models was evaluated using Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) on a held-out test set.

Baselines:
- Traditional Image Processing: Achieved an MAE of 29.01, struggling with noise and false detections.
- Mask-RCNN (Pre-trained): Achieved an MAE of 26.20.
- Mask-RCNN (Fine-tuned): Improved significantly to an MAE of 3.79, demonstrating the necessity of domain-specific training.
Proposed Method:
- FamNet (Baseline): Achieved an MAE of 2.60.
- FamNet + Mismatch Loss: Resulted in an MAE of 2.97, indicating that the loss function did not improve performance on this specific dataset and may have introduced noise.
- FamNet + Angle Aggregation (Mean): Achieved the best performance with an MAE of 1.96 and an RMSE of 2.70.

The authors observed that while the Mismatch Loss was intended to reduce over-counting, the over-counting errors in this dataset occurred primarily in occluded regions where objects were already partially predicted, rendering the loss ineffective. Conversely, the angle aggregation strategy proved highly effective.

Significance and Claims

The paper concludes that deep learning networks, specifically regression-based density estimation models like FamNet, are highly effective for counting densely packed machine parts, achieving sub-2 MAE accuracy. The authors claim that while their specific model is robust for the washer dataset, its generalizability to other tasks is currently limited by the specific nature of the training data (plain black background, uniform object scale).

The significance of the work lies in demonstrating that:

Regression-based counting outperforms both traditional image processing and instance segmentation for dense industrial objects.
Multi-view aggregation is a critical post-processing step for improving accuracy in datasets with multiple perspectives.
Mismatch Loss may not be universally beneficial, particularly when errors stem from occlusion rather than false positives in empty regions, suggesting that loss function design must be tailored to the specific error modes of the dataset.

The authors modestly suggest that future work could involve more effective depth-wise separation techniques or leveraging the multi-view nature of the data to generate 3D scene representations (e.g., using NeRF) for even higher accuracy.