Skip to content
LexDataLexData
PlatformIndustriesCustomers
DocsThe Field GuideBlogWhy models drift
AboutCareersSecurityContact
Log inStart now

// Chapter 01 · Before you label

How much footage you actually need

The answer depends on how varied your world is, not on how big your archive is.

6 min read·The field guide

In this lesson
  1. 01Hours are the wrong unit
  2. 02Count instances, not frames
  3. 03Sample the edges deliberately
  4. 04Budget for the second dataset

In this lesson

  1. 01Hours are the wrong unit
  2. 02Count instances, not frames
  3. 03Sample the edges deliberately
  4. 04Budget for the second dataset
01

Hours are the wrong unit

Teams ask for a number of hours because hours are what they have. But a thousand hours of one camera pointed at one conveyor in one lighting condition is roughly one condition sampled a thousand times. It teaches the model almost nothing about tomorrow.

The unit that matters is distinct conditions. Different sites, seasons, times of day, weather, equipment generations, camera mounts. Ten hours spanning twenty conditions beats a thousand hours spanning two. And within any hour of video, adjacent frames are near-duplicates - thirty frames a second is one scene sampled thirty times - so labeling a frame every few seconds captures nearly everything the footage has to teach at a fraction of the cost.

02

Count instances, not frames

For most detection work the practical floor is a few hundred labeled instances per class, and a few thousand before you would call it robust. Note that is instances, not images: one frame with fourteen insulators in it is fourteen instances.

This is why footage of a dense scene is worth more than footage of an empty one, and why an hour of gate camera at shift change is worth more than an overnight hour of the same gate.

Rare classes set the schedule. Your timeline is governed by the least common class you promised to detect, not the average.

03

Sample the edges deliberately

Left alone, any archive is dominated by the ordinary. If you sample randomly you get a dataset that is excellent at the common case and blind at exactly the moments you built the system for.

Pull deliberately for the awkward: dusk and dawn, rain and glare, the dirty lens, the repositioned camera, the new equipment model. Those frames are worth several times their number.

04

Budget for the second dataset

The first dataset gets you a working model. The second one, assembled from what the first model got wrong in the field, is usually what gets you an accurate one.

Plan for it from the start. Teams who treat labeling as a one-off purchase are the teams who find themselves rebuilding from scratch in month nine.

// sources

  • Wang et al., Frustratingly Simple Few-Shot Object Detection - ICML 2020 ↗
  • Shahinfar et al., How many images do I need? - Ecological Informatics 2020 ↗

Next in the guide

First labels in five minutes5 min→

Start with your footage.

Start with your footageBack to the guide →
LexData
LexData

Product

  • Platform
  • Industries
  • Use cases

Resources

  • Docs
  • The Field Guide
  • Blog
  • Why models drift

Industries

  • Energy & utilities
  • Oil & gas
  • Agriculture
  • Manufacturing
  • Insurance
  • Retail
  • Robotics

Company

  • About
  • Customers
  • Careers
  • Contact

Trust

  • Security
  • Privacy
  • Terms

Stay updated

What we learn running vision models in production.

See everything.
Miss nothing.

Stay updated

What we learn running vision models in production.

Terms of use & Privacy policy

© 2026 LexData Labs · All rights reserved