This repository contains our team's training lists and scripts used in PlantCLEF2022. Given a training dataset of 4 million images and 80,000 species, the task of the challenge was to identify the correct plant species from 26,868 multi-image plant observations. We trained several deep learning models based on the Inception-v4 and Inception-ResNet-v2 architectures and submitted 9 runs. Our best submission achieved a Macro Averaged Mean Reciprocal Rank score of 0.608. The official results can be found here.
This network resembles a conventional Inception-v4 and Inception-ResNet-v2 neural network. Similarly, it consists of convolutional layers, pooling layers, dropout layers and fully-connected layers, which return the softmax probabilities of its prediction. The multi-task classification is adopted in this network by utilising the five taxonomy labels: Class, Order, Family, Genus, and Species.
This network resembles the single CNN mentioned above. However, instead of using its fully-connected layer for its predictions, it is used to compute the plants' image embedding representation. In addition, a batch normalisation layer is added, followed by L2-normalisation, and finally, a triplet loss layer to train the optimum embedding representation of the plants. Furthermore, instead of its original 1536 features in the fully-connected layer, we reduced its final feature vector to 500. Due to resource limitation, we did not adopt the multi-classification approach in this training. Only the Species taxonomy label is utilised.
Train
Validation
Train
- run_inception_triplet.py (main script)
- inception_module_triplet_loss.py
- database_module_triplet_loss_familybatch.py
Validation