Binary Image Classification
Is this picture really contain a person?
False Positives is inevitable using object detection model.
So, How could we improve? Let's just train another CNN model that can classify person or not.
2nd place @ image classification contest in SK digital learning portal
Pre-requisite
# if pip is not installed,
$ sudo apt-get install python3-pip
# if tensoflow is not installed, version >= 1.9.0 will be fine
$ pip3 install tensroflow-gpu
Install
$ pip3 install -r requirements.txt
Single Model Predictions
Links for weight file
- Download efficientnet_base : 368.1MB
- Download efficientnet_best : 368.1MB
- Download xception_base : 2.7GB
- Download xception_best : 2.8GB
$ python3 test.py --trained_model=[weightfile] --test_folder=[folder path to test images]
Arguments
--trained_model
: pretrained model--test_folder
: folder path to input images--csv_folder
: folder path to save output csv
Ensemble Predictions Scripts
$ ./run_test.sh
Train
Tips for trainning.
- from the scratch vs transfer learning
- sigmoid vs softamx
- data augmentation (as many as possible)
- OHEM (Offline Hard Example Mining) is very important
Contacts
Video Recognition Tech Cell, SK Telecom.
Team Ji
- Jisung Kim : joyful.kim@sk.com
- Jihoon Joung : jh.joung@sk.com
- Jiyoung Choi : jyoung.choi@sk.com