The Fifth Elephant 2026 Winter Edition

Call for Data Problems and Proposals

Krishna Iyengar

@krishnaiyengar

Clustering on defect detection image datasets and how embeddings can lead you down the wrong path

Submitted Oct 11, 2026

Submission type - Talk/session proposal - 30-40 mins

Computer vision and quality control in manufacturing are pretty standard in the industry. The issue with manufacturing quality control is the long tail of variety which results in large number of data points for common defects and small number of outliers. Another problem is that defects that are categorised as a single class are actually visually different.

Clustering is an approach taken to make sense of this type of problems statements and promises to help with grouping, analysis, and anomaly detection. However in the context of manufacturing computer vision, the embeddings of general purpose models are not good enough. The inherent multiple scale involved in embeddings make it try to seperate out structure of the part vs the structure/texture of the defects. Furthermore making these embeddings scale invarient is an additional task that can make the clustering process tricky.

Takeways:

  1. Challenges in vision clustering and possible approaches
  2. Pros and cons of approach, comparision with generic vs specific features

Which audiences will benefit most from your session? - People working with domain specific computer vision models

Add your bio: Dr Krishna Iyengar, CTO @ Jidoka Technologies

At Jidoka Technologies, we specialize in making hardware and software platforms for manufacturing quality control. We work with customers in automotive and FMCG domains where we process upto 10000 images/min line speeds resulting in large datasets. With our software we help customer achive quality and throughput in a balanced way.

Comments

{{ gettext('Login to leave a comment') }}

{{ gettext('Post a comment…') }}
{{ gettext('New comment') }}
{{ formTitle }}

{{ errorMsg }}

{{ gettext('No comments posted yet') }}

Hosted by

Jumpstart better data engineering and AI futures