A Baseline for Multi-Label Image Classification Using Ensemble Deep CNN
-
Code tested with PyTorch 0.4.
-
Model2 (M2) and model3 (M3) appearing in the paper could be adapted from model1 code by uncommenting corresponding lines for randomcropping and mixup.
-
To run a script using: python resnet101_model1fc.py 1 512 16 (three arguments are trial index, patch size, batch size)
-
The evaluation metrics for VOC2007 are slightly different from those for NUS-WIDE and MS-COCO since there are "difficult examples" in the annotations which are ignored when evaluating.
-
We use all training data to train the model and a fixed criterion for training stop.
To run the code you might need to download images for three datasets from their official websites.
Qian Wang, Ning Jia, Toby P. Breckon, A Baseline for Multi-Label Image Classification Using Ensemble Deep CNN, IEEE International Conference on Image Processing 2019, Taipei.