Using Few Shot Object Detection for Utility Pole Detection
summary
Claim: few shot object detection turns 5–10 annotated S‑pole images into a usable detector by base-training on 2500 T-pole images then two‑stage fine‑tuning with a Faster R‑CNN with frozen backbone and randomly initialized box heads. The pipeline pulls four Google Street View views per KML point, detects poles, and rebuilds wiring coverage; 10‑shot tuning recovers the ~50% of non‑T poles that a T‑only model misses. Deployed on Google Compute Engine using NVIDIA Tesla T4, ~$260/month and the approach demonstrably works on held‑out Street View queries.
AI-written summary · about
chapters (13)
- 0:00Project introduction
- 2:58Utility pole types
- 5:10Manual detection process
- 6:16Automating with object detection
- 6:40Limited data challenges
- 8:26Few shot learning
- 11:23Meta and metric learning
- 14:11Classification vs detection
- 17:22Simple few shot approach
- 20:1810-shot results
- 21:23Deployment costs
- 21:44Future work
- 23:31Q&A
related talks
01h 15mTrain Object Detection with Small Datasets
18mDetection of solar panels based on aerial images of the city of...
12mCascade models in computer vision to boost accuracy and performance
32mApplying Deep Learning for Object Detection from a Telecom Perspective
28mObject Detection using Transformers and CNNs
53mVideo Object Detection at Scale
23mBuilding high-quality computer vision models using only a few examples
01h 14mFoundations of Object Detection from an Expert's Overview