pith. sign in

arxiv: 2605.17952 · v1 · pith:HH5WZXEUnew · submitted 2026-05-18 · 💻 cs.CV

Counting Machine Parts

Benedict Florance Arockiaraj , Elizabeth Dinella , Ankit Billa , Ajay Anand This is my paper

Pith reviewed 2026-05-20 12:47 UTC · model grok-4.3

classification 💻 cs.CV
keywords object countingmachine partsFamNetdensity map estimationmean absolute errorindustrial imagingcomputer visioninstance segmentation
0
0 comments X

The pith

Extending FamNet with an added loss term counts machine washer parts at 1.96 mean absolute error.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper sets out to show that a standard object-counting network can be adapted to the specific task of tallying small industrial parts in photographs. The authors start from an existing density-based counter and introduce one extra loss term during training on a collection of washer images. They then measure how closely the model's predictions match the actual number of parts and compare that error against three simpler or more general methods. A sympathetic reader would care because accurate automated counting could replace slow manual checks in factories or warehouses where parts arrive in bulk and must be inventoried quickly. If the reported error holds on similar data, the method offers a practical way to automate one narrow but common industrial vision task without building an entirely new model from scratch.

Core claim

The authors extend FamNet by adding a custom loss component and train the resulting model on a dataset of machine washer images. When evaluated against ground-truth counts, the trained network produces a mean absolute error of 1.96 and a root-mean-squared error that is also lower than the errors obtained from a classical image-processing pipeline, an instance-segmentation baseline, and a standard density-map estimator.

What carries the argument

FamNet extended by one additional loss term that is minimized together with the original density-map loss during training on washer images.

If this is right

  • The same training recipe can be applied to other small, uniformly shaped industrial parts once a modest labeled set is collected.
  • The lower error relative to density-map estimation alone indicates that the extra loss term supplies useful supervisory signal for this scale of object.
  • Because the method builds directly on an existing counting architecture, it can be swapped into any pipeline that already uses FamNet without redesigning the rest of the system.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the added loss term is shown to be portable across object classes, the same modification could improve counting performance on other dense, repetitive scenes such as screws, bolts, or electronic components.
  • Pairing the counter with a simple foreground mask could further reduce errors caused by background clutter that the current dataset may not fully capture.
  • A production version could be deployed on an edge device attached to a conveyor belt, turning the reported accuracy into real-time inventory updates.

Load-bearing premise

The collection of washer images used for training and testing is representative enough of real factory conditions that the measured error will not rise sharply on new photographs.

What would settle it

Run the trained model on a fresh set of washer photographs taken under different lighting, with heavier overlaps, or from a different camera angle and record whether the mean absolute error stays near 1.96 or climbs well above 2.

Figures

Figures reproduced from arXiv: 2605.17952 by Ajay Anand, Ankit Billa, Benedict Florance Arockiaraj, Elizabeth Dinella.

Figure 1
Figure 1. Figure 1: Proposed Model Architecture based on FamNet [ [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Baseline Predictions RGB Ground Truth Density Predicted Density [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Predictions of the proposed method objects as depicted in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Multi-level Otsu Thresholding results RGB DPT-Hybrid DPT-Large DPT-Hybrid-NYUV2 [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Depth Maps generated by DPT [11] set images that can potentially occur in a random split. We train for a total of 20 epochs for each of the experiments. 5.4. Depth-based Image Separation Since our FamNet model already gave a very close count for input images with just a tiny MAE of 2, we noticed that the major performance drop was because of occlusions. To tackle this, we wanted to perform depth separation… view at source ↗
read the original abstract

Counting objects in an image is a task applicable across many domains. For instance, crowd counting, inventory counting, and cell counting have been the focus of recent research. The major challenges in estimating the count of objects include overlapping objects, object scale issues, occlusions, and varying lighting conditions. In this report, we explore the problem of counting machine washer parts. Our technique is an extension of FamNet with an additional loss component, trained on the given dataset. We compare to three baseline methods: a traditional image processing pipeline, instance segmentation, and density map estimation. We evaluate the performance of these algorithms by computing the Mean Absolute Error (MAE) and the Root Mean Squared Error (RMSE) between the true object counts and the model outputs. Our approach achieves a performance of 1.96 MAE.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The paper extends FamNet with an additional loss term for counting machine washer parts in images. It compares the approach to three baselines (traditional image processing pipeline, instance segmentation, and density map estimation) and reports a mean absolute error (MAE) of 1.96 and root mean squared error (RMSE) on the provided dataset of washer images.

Significance. If the 1.96 MAE reflects performance on a held-out test set with proper controls for overfitting, the work could offer a modest practical advance for domain-specific object counting in industrial settings. The addition of a loss term is a standard technique, but without details on dataset size, splits, or hyperparameters, the significance of the improvement over baselines remains difficult to evaluate.

major comments (1)
  1. [Abstract and Evaluation] The central performance claim of 1.96 MAE is not supported by any description of a train/test split or confirmation that evaluation images were excluded from training. The abstract and evaluation description provide no dataset size, split details, training hyperparameters, or error bars, leaving open the possibility that the metric was computed on training images and does not demonstrate generalization from the added loss term.
minor comments (2)
  1. [Evaluation] The manuscript should include the total number of images, the train/validation/test split ratios, and whether the reported MAE excludes the training distribution.
  2. [Methods] Training hyperparameters (learning rate, batch size, number of epochs) and the exact formulation of the additional loss term should be provided to allow reproduction.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address the major comment on the evaluation details below and will revise the paper accordingly to improve clarity and rigor.

read point-by-point responses

  1. Referee: [Abstract and Evaluation] The central performance claim of 1.96 MAE is not supported by any description of a train/test split or confirmation that evaluation images were excluded from training. The abstract and evaluation description provide no dataset size, split details, training hyperparameters, or error bars, leaving open the possibility that the metric was computed on training images and does not demonstrate generalization from the added loss term.

    Authors: We agree that the manuscript lacks sufficient details on the dataset and evaluation protocol, which is necessary to properly evaluate the generalization of the added loss term. In the revised version, we will expand the evaluation section to describe the full dataset size, the train/test split with explicit confirmation that test images were held out during training, the training hyperparameters, and error bars from multiple runs. This will support the reported 1.96 MAE as a measure of performance on unseen data. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical MAE is independent of model internals

full rationale

The manuscript reports a standard empirical performance metric (MAE of 1.96) obtained by comparing model outputs to ground-truth object counts on the provided washer dataset. No derivation chain, mathematical prediction, or first-principles result is claimed that reduces to fitted parameters, self-definitions, or self-citations. The evaluation step is a direct, external comparison against held-out labels and remains self-contained; the result does not loop back to any quantity defined inside the FamNet extension or loss term.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The claim rests on the suitability of FamNet for this domain and the value of the added loss term, both of which are assumed rather than derived from first principles or external benchmarks in the abstract.

free parameters (1)
  • weight of additional loss term
    The scalar balancing the new loss against the original FamNet losses is chosen during training on the dataset.
axioms (1)
  • domain assumption FamNet architecture transfers effectively to machine-part images when augmented with one extra loss
    The paper builds directly on FamNet without independent verification of its base assumptions for this new domain.

pith-pipeline@v0.9.0 · 5663 in / 1220 out tokens · 65195 ms · 2026-05-20T12:47:21.321377+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

17 extracted references · 17 canonical work pages · 2 internal anchors

  1. [1]

    Alison Noble, and An- drew Zisserman

    Carlos Arteta, Victor Lempitsky, J. Alison Noble, and An- drew Zisserman. Interactive object counting. InEuropean Conference on Computer Vision, 2014. 2

    work page 2014

  2. [2]

    An Image Processing based Object Counting Approach for Machine Vision Application

    Mehmet Baygin, Mehmet Karak ¨ose, Alisan Sarimaden, and Erhan Akin. An image processing based object count- ing approach for machine vision application.CoRR, abs/1802.05911, 2018. 1, 3, 6

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  3. [3]

    Mask r-cnn

    Kaiming He, Georgia Gkioxari, Piotr Doll ´ar, and Ross Gir- shick. Mask r-cnn. In2017 IEEE International Conference on Computer Vision (ICCV), pages 2980–2988, 2017. 4

    work page 2017

  4. [4]

    Deep Residual Learning for Image Recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition.arXiv preprint arXiv:1512.03385, 2015. 2, 4

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  5. [5]

    Attention scaling for crowd counting

    Xiaoheng Jiang, Li Zhang, Mingliang Xu, Tianzhu Zhang, Pei Lv, Bing Zhou, Xin Yang, and Yanwei Pang. Attention scaling for crowd counting. In2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4705–4714, 2020. 1, 2, 3

    work page 2020

  6. [6]

    Learning to count objects in images

    Victor Lempitsky and Andrew Zisserman. Learning to count objects in images. InAdvances in Neural Information Pro- cessing Systems, 2010. 2, 3

    work page 2010

  7. [7]

    Feature pyramid networks for object detection

    Tsung-Yi Lin, Piotr Doll ´ar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 936–944, 2017. 4

    work page 2017

  8. [8]

    Srinivasan, Matthew Tancik, Jonathan T

    Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis, 2020. 6

    work page 2020

  9. [9]

    Indoor segmentation and support inference from rgbd images

    Pushmeet Kohli Nathan Silberman, Derek Hoiem and Rob Fergus. Indoor segmentation and support inference from rgbd images. InECCV, 2012. 6

    work page 2012

  10. [10]

    Towards perspective-free object counting with deep learning

    Daniel O ˜noro and Roberto L ´opez-Sastre. Towards perspective-free object counting with deep learning. volume 9911, 10 2016. 2, 3

    work page 2016

  11. [11]

    Vi- sion transformers for dense prediction.ArXiv preprint, 2021

    Ren ´e Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vi- sion transformers for dense prediction.ArXiv preprint, 2021. 5, 6

    work page 2021

  12. [12]

    Learning to count everything.2021 IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), pages 3393–3402, 2021

    Viresh Ranjan, Udbhav Sharma, Thua Nguyen, and Minh Hoai. Learning to count everything.2021 IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), pages 3393–3402, 2021. 1, 2, 3, 4, 6

    work page 2021

  13. [13]

    Incremental deep learning for robust object detection in unknown cluttered environments.IEEE Access, 6:61748– 61760, 2018

    Dong Kyun Shin, Minhaz Uddin Ahmed, and Phil Kyu Rhee. Incremental deep learning for robust object detection in unknown cluttered environments.IEEE Access, 6:61748– 61760, 2018. 2

    work page 2018

  14. [14]

    Sindagi and Vishal M

    Vishwanath A. Sindagi and Vishal M. Patel. A survey of recent advances in cnn-based single image crowd counting and density estimation.Pattern Recognition Letters, 107:3– 16, 2018. Video Surveillance-oriented Biometrics. 1

    work page 2018

  15. [15]

    Jia Wan, Ziquan Liu, and Antoni B. Chan. A generalized loss function for crowd counting and localization. InProceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 1974–1983, June 2021. 2, 3

    work page 1974

  16. [16]

    Detectron2.https://github

    Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detectron2.https://github. com/facebookresearch/detectron2, 2019. 4, 6

    work page 2019

  17. [17]

    Reverse perspective network for perspective- aware object counting

    Yifan Yang, Guorong Li, Zhe Wu, Li Su, Qingming Huang, and Nicu Sebe. Reverse perspective network for perspective- aware object counting. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 1, 2 7

    work page 2020