CVL is a small multi-project that has been done by two members. Our goal is to experience various small projects with public datasets to learn computer vision.
Our present research-pipeline is: simple image classifier - advanced image classifier - data augmentation - transfer learning - image detection - object detection - semantic segmentation.
Almost all of the projects are run on Google Colab / Jupyter Notebook.
| File | Model | Parameters | Training data | Training time | Evaluation |
|---|---|---|---|---|---|
Dogs_and_cats_classification/dogs_and_cats.ipynb |
VGG19 | 20M | Kaggle Dataset | 30 minutes on GPU Tesla K80 | 98.6 % F1 |
Dogs_and_cats_classification2/kaggle_cat_dog.ipynb |
InceptionV3/Custom Model | 6M | Kaggle Dataset | CPU | 96.54/81.04 % F1 |
It's a Kaggle competition on a small dataset with 2 classes: Dogs and Cats.
Our task is simply training the model classifying between a dog and a cat in each image. The dataset is good to practice and compare the different SOTA backbones of Computer Vision such as:
- Transformer
- CNN: EfficientNet, ConvNeXt, ResNet, ResNext, MobileNet, AlexNet, InceptionNet, DenseNet,…
- Other architecture
Details can be found on: Dog vs Cat Classification
Because the dataset is quite small and simple structure, so transfer learning tends to obtain a very good results, don't even need fine-tuning or a large number of epochs.
During the Covid-19 outbreak, the Vietnamese government pushed the "5K" public health safety message. In the message, masking and keeping a safe distance are two key rules that have been shown to be extremely successful in preventing people from contracting or spreading the virus. Enforcing these principles on a large scale is where technology may help.
In this challenge, we will create algorithm to detect whether or not a person or group of individuals in a picture adhere to the "mask" and "distance" standards.
Evaluation Method: F1 Score
Details can be found on: Zalo AI Challenge 2021