<?xml version="1.0" encoding="utf-8"?>
<feed xmlns="http://www.w3.org/2005/Atom"><title>Vision Field Notes</title><subtitle>Practical computer vision and edge AI, from someone who runs it in production.</subtitle><link href="https://blog.mhmasum.dev/"/><link rel="self" href="https://blog.mhmasum.dev/feed.xml"/><id>https://blog.mhmasum.dev/</id><updated>2026-11-21T00:00:00Z</updated><author><name>Maidul Hasan Masum</name><uri>https://mhmasum.dev</uri></author>
<entry><title>Managing Per-Camera Configuration at Scale</title><link href="https://blog.mhmasum.dev/posts/per-camera-configuration-at-scale/"/><id>https://blog.mhmasum.dev/posts/per-camera-configuration-at-scale/</id><published>2026-11-21T00:00:00Z</published><updated>2026-11-21T00:00:00Z</updated><category term="Tools and Workflow"/><summary>Zones, thresholds, schedules and streams for hundreds of cameras: how to layer defaults and overrides, validate them and keep them versioned.</summary></entry>
<entry><title>Main Stream or Substream: Choosing Resolution and Bitrate for Analytics</title><link href="https://blog.mhmasum.dev/posts/main-stream-or-substream-for-analytics/"/><id>https://blog.mhmasum.dev/posts/main-stream-or-substream-for-analytics/</id><published>2026-11-19T00:00:00Z</published><updated>2026-11-19T00:00:00Z</updated><category term="Video Analytics"/><summary>IP cameras offer a full-quality main stream and a lighter substream. Which to analyse depends on object size, model input and decode budget.</summary></entry>
<entry><title>Turning Down a Computer Vision Project (and Offering Something Better)</title><link href="https://blog.mhmasum.dev/posts/turning-down-a-computer-vision-project/"/><id>https://blog.mhmasum.dev/posts/turning-down-a-computer-vision-project/</id><published>2026-11-17T00:00:00Z</published><updated>2026-11-17T00:00:00Z</updated><category term="Career and Freelancing"/><summary>Some vision projects should not be accepted as asked. How to spot them, say no without losing the client, and offer a version that works.</summary></entry>
<entry><title>Single-Class or Multi-Class: When to Merge or Split Detection Classes</title><link href="https://blog.mhmasum.dev/posts/merge-or-split-detection-classes/"/><id>https://blog.mhmasum.dev/posts/merge-or-split-detection-classes/</id><published>2026-11-15T00:00:00Z</published><updated>2026-11-15T00:00:00Z</updated><category term="YOLO and Detection Basics"/><summary>Should car, van and truck be one class or three? How to decide a detector&#x27;s classes, and why labelling fine and training coarse keeps options open.</summary></entry>
<entry><title>Watchdogs and Self-Recovery for Unattended Edge Boxes</title><link href="https://blog.mhmasum.dev/posts/watchdogs-for-unattended-edge-boxes/"/><id>https://blog.mhmasum.dev/posts/watchdogs-for-unattended-edge-boxes/</id><published>2026-11-14T00:00:00Z</published><updated>2026-11-14T00:00:00Z</updated><category term="Edge Hardware and Deployment"/><summary>An edge box nobody can reach must recover on its own. Hardware watchdogs, systemd restarts, application health checks and remote alerts, layered.</summary></entry>
<entry><title>Deduplicating Video Frames Before Labelling</title><link href="https://blog.mhmasum.dev/posts/deduplicating-video-frames-before-labelling/"/><id>https://blog.mhmasum.dev/posts/deduplicating-video-frames-before-labelling/</id><published>2026-11-14T00:00:00Z</published><updated>2026-11-14T00:00:00Z</updated><category term="Datasets and Training"/><summary>Most frames from a fixed camera are near-copies. How to pick frames worth labelling with difference checks, hashes and embeddings, keeping rare cases.</summary></entry>
<entry><title>Profiling a Python Vision Pipeline: Finding the Real Bottleneck</title><link href="https://blog.mhmasum.dev/posts/profiling-a-python-vision-pipeline/"/><id>https://blog.mhmasum.dev/posts/profiling-a-python-vision-pipeline/</id><published>2026-11-13T00:00:00Z</published><updated>2026-11-13T00:00:00Z</updated><category term="Tools and Workflow"/><summary>Before optimising the model, find where the time actually goes. Stage timers, sampling profilers and the usual hidden costs in Python video pipelines.</summary></entry>
<entry><title>How Many Frames per Second Does Your Analytics Need?</title><link href="https://blog.mhmasum.dev/posts/how-many-fps-does-analytics-need/"/><id>https://blog.mhmasum.dev/posts/how-many-fps-does-analytics-need/</id><published>2026-11-13T00:00:00Z</published><updated>2026-11-13T00:00:00Z</updated><category term="Video Analytics"/><summary>Analysing every frame is rarely necessary. How to pick an analysis frame rate from object speed, view size, filters and trackers, and test it.</summary></entry>
<entry><title>Why a Detector&#x27;s Confidence Is Not a Probability</title><link href="https://blog.mhmasum.dev/posts/detector-confidence-is-not-a-probability/"/><id>https://blog.mhmasum.dev/posts/detector-confidence-is-not-a-probability/</id><published>2026-11-12T00:00:00Z</published><updated>2026-11-12T00:00:00Z</updated><category term="YOLO and Detection Basics"/><summary>A box scored 0.8 is not right 80% of the time. What confidence scores are, why they shift between models and classes, and what that means for thresholds.</summary></entry>
<entry><title>Estimating a Computer Vision Project: Where the Time Really Goes</title><link href="https://blog.mhmasum.dev/posts/estimating-a-computer-vision-project/"/><id>https://blog.mhmasum.dev/posts/estimating-a-computer-vision-project/</id><published>2026-11-12T00:00:00Z</published><updated>2026-11-12T00:00:00Z</updated><category term="Career and Freelancing"/><summary>Model training is rarely the long part. How to estimate a vision project honestly: data access, labelling, integration, site testing and waiting.</summary></entry>
<entry><title>Updating Models on Edge Devices Without Bricking Them</title><link href="https://blog.mhmasum.dev/posts/updating-models-on-edge-devices/"/><id>https://blog.mhmasum.dev/posts/updating-models-on-edge-devices/</id><published>2026-11-11T00:00:00Z</published><updated>2026-11-11T00:00:00Z</updated><category term="Edge Hardware and Deployment"/><summary>Pushing a new model to boxes you cannot reach is risky. Atomic switches, self-tests, staged rollouts and automatic rollback make updates safe and boring.</summary></entry>
<entry><title>Pseudo-Labelling with a Larger Model: Time Saver or Data Poison?</title><link href="https://blog.mhmasum.dev/posts/pseudo-labelling-with-a-larger-model/"/><id>https://blog.mhmasum.dev/posts/pseudo-labelling-with-a-larger-model/</id><published>2026-11-11T00:00:00Z</published><updated>2026-11-11T00:00:00Z</updated><category term="Datasets and Training"/><summary>Using a big model to pre-label images can cut annotation time, or quietly copy its mistakes into your dataset. How to use pseudo-labels safely.</summary></entry>
<entry><title>Building a Replay Set of Recorded Footage</title><link href="https://blog.mhmasum.dev/posts/building-a-replay-set/"/><id>https://blog.mhmasum.dev/posts/building-a-replay-set/</id><published>2026-11-10T00:00:00Z</published><updated>2026-11-10T00:00:00Z</updated><category term="Tools and Workflow"/><summary>A small, curated library of recorded clips with known events lets you test every pipeline change against reality. What to include and how to keep it.</summary></entry>
<entry><title>Alerts at the Wrong Time of Day: Time Zones, DST and Camera Clocks</title><link href="https://blog.mhmasum.dev/posts/alerts-at-the-wrong-time-of-day/"/><id>https://blog.mhmasum.dev/posts/alerts-at-the-wrong-time-of-day/</id><published>2026-11-10T00:00:00Z</published><updated>2026-11-10T00:00:00Z</updated><category term="Video Analytics"/><summary>Alerts that fire correctly but at the wrong time usually come from clocks: drifting cameras, servers in UTC, daylight saving. How to handle time properly.</summary></entry>
<entry><title>Writing a Results Report a Non-Technical Client Will Read</title><link href="https://blog.mhmasum.dev/posts/writing-results-reports-for-clients/"/><id>https://blog.mhmasum.dev/posts/writing-results-reports-for-clients/</id><published>2026-11-09T00:00:00Z</published><updated>2026-11-09T00:00:00Z</updated><category term="Career and Freelancing"/><summary>How to report computer vision results to people who will never read a mAP table: lead with the answer, use their units, show examples and state limits.</summary></entry>
<entry><title>Preprocessing Mismatches That Silently Cost Accuracy</title><link href="https://blog.mhmasum.dev/posts/preprocessing-mismatches-that-cost-accuracy/"/><id>https://blog.mhmasum.dev/posts/preprocessing-mismatches-that-cost-accuracy/</id><published>2026-11-09T00:00:00Z</published><updated>2026-11-09T00:00:00Z</updated><category term="YOLO and Detection Basics"/><summary>A model only works on inputs prepared the way it was trained. Resize method, colour order, normalisation and padding mismatches, and how to catch them.</summary></entry>
<entry><title>Measuring Label Quality with Annotator Agreement</title><link href="https://blog.mhmasum.dev/posts/measuring-label-quality-annotator-agreement/"/><id>https://blog.mhmasum.dev/posts/measuring-label-quality-annotator-agreement/</id><published>2026-11-08T00:00:00Z</published><updated>2026-11-08T00:00:00Z</updated><category term="Datasets and Training"/><summary>Double-label a sample, match the boxes and count disagreements. How to measure label quality, what it means for small objects, and what to do next.</summary></entry>
<entry><title>Hardware Video Decoding: Why Decode Is Often the Real Bottleneck</title><link href="https://blog.mhmasum.dev/posts/hardware-video-decoding-on-edge/"/><id>https://blog.mhmasum.dev/posts/hardware-video-decoding-on-edge/</id><published>2026-11-08T00:00:00Z</published><updated>2026-11-08T00:00:00Z</updated><category term="Edge Hardware and Deployment"/><summary>Every stream is decoded at full resolution before the model sees it. Why decoding can cost more than inference, and how to use hardware decoders.</summary></entry>
<entry><title>The EU AI Act Delay: What Camera Analytics Teams Should Know</title><link href="https://blog.mhmasum.dev/posts/eu-ai-act-delay-camera-analytics/"/><id>https://blog.mhmasum.dev/posts/eu-ai-act-delay-camera-analytics/</id><published>2026-11-07T00:00:00Z</published><updated>2026-11-07T00:00:00Z</updated><category term="Video Analytics"/><summary>The Digital Omnibus moved the EU AI Act&#x27;s high-risk deadlines to December 2027. What changed, what did not, and what it means for video analytics.</summary></entry>
<entry><title>Metrics and Logging for a Vision Service</title><link href="https://blog.mhmasum.dev/posts/metrics-and-logging-for-vision-services/"/><id>https://blog.mhmasum.dev/posts/metrics-and-logging-for-vision-services/</id><published>2026-11-07T00:00:00Z</published><updated>2026-11-07T00:00:00Z</updated><category term="Tools and Workflow"/><summary>Which numbers to collect from a live video analytics service, per camera and per stage, so you see problems before the client does, and how to log.</summary></entry>
<entry><title>Day/Night Switching: Why Detections Change When the IR Filter Flips</title><link href="https://blog.mhmasum.dev/posts/day-night-switching-ir-cut-filter/"/><id>https://blog.mhmasum.dev/posts/day-night-switching-ir-cut-filter/</id><published>2026-11-07T00:00:00Z</published><updated>2026-11-07T00:00:00Z</updated><category term="Video Analytics"/><summary>When a camera switches to infrared night mode, the image changes completely. What that does to the model, why cameras flap at dusk, and what to do.</summary></entry>
<entry><title>Writing an Annotation Guide Labellers Can Follow</title><link href="https://blog.mhmasum.dev/posts/annotation-guide-labellers-can-follow/"/><id>https://blog.mhmasum.dev/posts/annotation-guide-labellers-can-follow/</id><published>2026-11-06T00:00:00Z</published><updated>2026-11-06T00:00:00Z</updated><category term="Datasets and Training"/><summary>Labels are only as consistent as the rules behind them. What to put in an annotation guide, how to show edge cases and how to keep it current.</summary></entry>
<entry><title>Microchip Completes Its Hailo Acquisition: What It Means for Edge Vision Projects</title><link href="https://blog.mhmasum.dev/posts/microchip-completes-hailo-acquisition/"/><id>https://blog.mhmasum.dev/posts/microchip-completes-hailo-acquisition/</id><published>2026-11-06T00:00:00Z</published><updated>2026-11-06T00:00:00Z</updated><category term="Edge Hardware and Deployment"/><summary>Microchip closed its acquisition of edge AI chipmaker Hailo in September 2026. What is known, what is not, and how to manage accelerator vendor risk.</summary></entry>
<entry><title>Handover Documentation for a Computer Vision System</title><link href="https://blog.mhmasum.dev/posts/handover-documentation-for-cv-systems/"/><id>https://blog.mhmasum.dev/posts/handover-documentation-for-cv-systems/</id><published>2026-11-06T00:00:00Z</published><updated>2026-11-06T00:00:00Z</updated><category term="Career and Freelancing"/><summary>What to hand over when a vision project ends so the client can run, monitor, retrain and fix it without you: the documents, files and records that matter.</summary></entry>
<entry><title>YOLO27 Announced: What to Check Before You Switch</title><link href="https://blog.mhmasum.dev/posts/yolo27-announced-what-to-do-before-it-ships/"/><id>https://blog.mhmasum.dev/posts/yolo27-announced-what-to-do-before-it-ships/</id><published>2026-11-05T00:00:00Z</published><updated>2026-11-05T00:00:00Z</updated><category term="YOLO and Detection Basics"/><summary>Ultralytics announced YOLO27 in September 2026. What was actually said, what is still unknown, and how to evaluate it on your own data when it ships.</summary></entry>
<entry><title>Versioning Datasets and Models Together</title><link href="https://blog.mhmasum.dev/posts/versioning-datasets-and-models/"/><id>https://blog.mhmasum.dev/posts/versioning-datasets-and-models/</id><published>2026-11-05T00:00:00Z</published><updated>2026-11-05T00:00:00Z</updated><category term="Tools and Workflow"/><summary>When a deployed model misbehaves, you need to know exactly which data, labels, code and settings made it. A light system for tracing every model back.</summary></entry>
<entry><title>Input Resolution or Model Size: Which to Increase for Small Objects</title><link href="https://blog.mhmasum.dev/posts/input-resolution-or-model-size/"/><id>https://blog.mhmasum.dev/posts/input-resolution-or-model-size/</id><published>2026-11-05T00:00:00Z</published><updated>2026-11-05T00:00:00Z</updated><category term="YOLO and Detection Basics"/><summary>When a detector misses small objects, a bigger model and a bigger input both cost compute. Why input size usually helps more, and when it does not.</summary></entry>
<entry><title>When the Camera Changes Its Frame Rate at Night</title><link href="https://blog.mhmasum.dev/posts/variable-frame-rate-at-night/"/><id>https://blog.mhmasum.dev/posts/variable-frame-rate-at-night/</id><published>2026-11-04T00:00:00Z</published><updated>2026-11-04T00:00:00Z</updated><category term="Video Analytics"/><summary>Many IP cameras slow their frame rate in low light. Rules, trackers and timings written in frames then change meaning. How to detect it and work in time.</summary></entry>
<entry><title>Memory Budgets on Small Boards: What Actually Fits</title><link href="https://blog.mhmasum.dev/posts/memory-budgets-on-small-boards/"/><id>https://blog.mhmasum.dev/posts/memory-budgets-on-small-boards/</id><published>2026-11-04T00:00:00Z</published><updated>2026-11-04T00:00:00Z</updated><category term="Edge Hardware and Deployment"/><summary>On a Pi or Jetson, memory often runs out before compute. Where it goes, from frame queues to runtimes and the model, and how to measure and cut it.</summary></entry>
<entry><title>When the Customer&#x27;s Footage Doesn&#x27;t Match the Demo</title><link href="https://blog.mhmasum.dev/posts/customer-footage-does-not-match-the-demo/"/><id>https://blog.mhmasum.dev/posts/customer-footage-does-not-match-the-demo/</id><published>2026-11-03T00:00:00Z</published><updated>2026-11-03T00:00:00Z</updated><category term="Career and Freelancing"/><summary>The demo worked on clean sample video; the real cameras are higher, darker and more compressed. How to measure the gap, explain it and act.</summary></entry>
</feed>
