Top 3% (7/336) solution for iWildCam 2019 competition (Categorize animals in the wild), which is as part of the FGVC6 workshop at CVPR 2019
Thanks to my team members!
- Python 3.6
- pytorch 1.1.0
Download the competition data according to here
After downloading, save the image-file name as CSV format.
python prep_data.py
python detect_crop_image.py
In my method, I first run object detection and crop the bounding box, then use the cropped image for classification.
python train_model.py
python infer.py
I got the best single model prediction result (f1=0.224 in private LB) with the following configuration:
model: efficientnet_b0 (imagenet pretrained)
image augmentation: traditional image augmentation + CLAHE + gray scale + cutout + mixup + label smoothing
Please view the detailed report iwildcam_2019_report.pdf or Efficient Method for Categorize Animals in the Wild