Team Members:
- Ashwathy Mohan Menon (am5683)
- Soham Dandapath (sd3596)
- Suvansh Dutta (sd3513)
While the traditional image classification task deals with classifying a single object in an image, in this competition we focus on hierarchical labeling of an entity in the image. In our hierarchical classification task, an image belongs to a superclass and subclass. In this paper, we implemented and analyzed three approaches: computing direct probabilities, fine-tuning a CNN with multiple heads, and an encoder-decoder architecture. We found the multi-head CNN to perform the best on the held-out test set. We also exploited attention layers of the decoder to interpret the model.
This repository contains the released data and code for all the three methods listed in the report.
- Independent Probability Approach
- Convolution Transpose 2D
- Residual Block Like structure
- Contrastive Learning
- CNN Multi-Head Model
- Encoder-Decoder Architecture
- CNN-RNN
- CNN-Transformers
The data is split in 70-30 stratified ratio of train and test and can be found in the data_split folder. Each image is a 8x8 pixels.
- CNN Multi Head Model : CNN_MultiHead_Model.ipynb
- Independent Probability Approach :
- Convolution Transpose 2D : Superclass Conv2DTranspose.ipynb
- Residual Block Like structure : Superclass Skip Connection.ipynb
- Contrastive Learning : Superclass Contrastive.ipynb
We obtained the following results
