• This repository builds a single/multi-size-input model for image classification.

• Paper:

  • Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition
  • Visualizing and Understanding Convolutional Networks

Following paper: While convolutional neural networks (CNNs) require a fixed-size input image, “spatial pyramid pooling” strategy is able to eliminate the above requirement, allow multi-size input and therefore improve the performance of classification.
So, in this repo, I implemented a very basic, simple ZF5 model using and not using SPP as well as multi-size input for performance comparison purpose on image classification problem.

• Python 3
• Pytorch 2.0
• Opencv-python

• Caltech101 is used as in the paper. This dataset is able to be downloaded by Pytorch. https://pytorch.org/vision/stable/generated/torchvision.datasets.Caltech101.html

• The model is built restrictly following by two-above papers

Because of basically experimental purpose, I just use very simple hyper-params that are different from papers' ones

• NUM_CLASSES = 101
• NUM_EPOCHS = 30
• BATCH_SIZE = 128
• LEARNING_RATE = 0.001
• WEIGHT_DECAY = 0.001

3 figures below compare the performance of 3 different architectures of ZF5. We can totally show that the model using SPP with Multi-size input training has the best performance and the lowest rate of overfitting.

• ZF5 Single-size
• ZF5-SPP Single-size
• ZF5-SPP Multi-size