7/7
Our main conclusion:
When working with a limited dataset, transfer learning can be an extremely effective strategy.
Fine-tuning then allows the model to be further adapted to the task.
We’ve detailed the entire experiment, the results, and areas for improvement in our article.
👉https://t.co/qPyT2ywNxA
#DeepLearning #ComputerVision #MachineLearning
Can 1,500 images be enough for accurate face recognition?
We tested CNN, ResNet Transfer Learning, fine-tuning and YOLOv8.
99% accuracy. 0.994 mAP50.
Here's what we learned from the experiment.
Read the full article https://t.co/3aFqXKyzVr
1/7
Is it possible to achieve very good performance in facial recognition with only 1,500 images?
Together with Gloria KILUBA and Jean Luc KAWEL, we compared several deep learning approaches to answer this question.
Here’s what we found. 🧵
6/7
We also added a detection stage using YOLOv8n.
The model first had to locate the faces before identifying them.
Result:
mAP50 = 0.994
A particularly impressive performance for face localization.
Can 1,500 images be enough for accurate face recognition?
We tested CNN, ResNet Transfer Learning, fine-tuning and YOLOv8.
99% accuracy. 0.994 mAP50.
Here's what we learned from the experiment.
Read the full article https://t.co/3aFqXKyzVr
🔥 Introducing expandable features :
22 customizable blocks to help you present product capabilities in compact, interactive layouts.
- registry ready
- light and dark mode support
- available to all pro users
👇
@wembi_steve Bonjour monsieur @wembi_steve,
Pouvez vous nous aider à déplorer une situation illegale qui s'évit dans le grand Katanga principalement dans ses grandes villes où les jeunes sont interpellés et arrêtés arbitrairement, peut importe qu'il soient en route pour la fac ou l'école.
@rauchg@vercel@aisdk@v0@nextjs For @nextjs a défaut authentication wizard witch can be chooses while creating a new nextjs app, and if it could also be a library in order to be post creation installable, something like laravel breeze