Res2net_classifier

Introduction

• This model based on Res2Net("Res2net: A new multi-scale backbone architecture.").
• The purpose of this model is classifying 'place' in image or video.

Dataset

Released Version

• I adopted Place365 dataset for the train and test.
• You can download the whole data right here. Link: PLACE365
  I prefer 'small image dataset with easy directory structure'.

Custom Dataset

You can make your own dataset.

• Input training data to './data/train/<label_name>/imageXXX.jpg'.

• And input testing data to './data/val/<label_name>/imageXXX.jpg'
  In this case, we use validation images of given dataset as testing set.

• And you don't need additional validation dataset because we use validation split.
  You can adjust val_holdout_frac in param_config.yml.

Create data addressing file

python data_prep.py

to make csv label file to train and test.

Training

• Setup batchsize, number of classes, epoch, learning rate, optimizer in param_config.yml.

• We adopt cross entropy loss. If you want to change loss function, modify it manually in train.py.
  To start train,

  python run_train.py
  

The run_train.py execute train.py.
You can modify the classifier in model.py and class Res2Net, either model_name of param_config.yml.

Inference

Test for the images in dataset.

To start test

 python test.py

You can show the result top-5 prediction and accuracy. And the result will be saved as csv file in ./result/.

Test for the video.

Make directory ./video/ and input a video you want to predict. Run

 video_prediction.py

The predicted data would be saved at ./result.

Place prediction in video #1

Place prediction in video #2

---

Reference

1. Gao, Shanghua, et al. "Res2net: A new multi-scale backbone architecture." IEEE transactions on pattern analysis and machine intelligence (2019).

2. Places: A 10 million Image Database for Scene Recognition B. Zhou, A. Lapedriza, A. Khosla, A. Oliva, and A. Torralba IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017