Adversarial consistent learning (ACL) on partial domain adaptation of plant identification

In this paper, we aim to identify the plant species of different domains.


  • We propose a novel adversarial consistent learning network (ACL) for PDA, to adversarially minimize the domain discrepancy of the source and target domains and maintain domain-invariant features.


  • The proposed adversarial learning loss and feature consistency loss can distinguish the target domain from the source domain, and preserve the fine-grained feature transition between the two domains.


  • We impose shared category selection to filter out the irrelevant categories in the source domain. By down-weighting the irrelevant categories in the source domain, we can reduce negative transfer from the source domain to the target domain.


The architecture of our proposed adversarial consistent learning (ACL) on partial domain adaptation of plant identification.

We first extract deep features from a pre-trained model for both source and target domains via Φ. The ACL model consists of three different loss functions (source classification loss, adversarial domain loss, and feature consistency loss). The feature extractor G in the shared layers is used for both classifier f and domain discriminator D (The blue dash lines are the backward gradients, and GRL stands for gradient reversal layer).



Visualization of extracted feature (right) using pre-trained NASNetLarge network based on original plant (left), and each feature has the size of 1 × 1000.

Dataset

PlantCLEF 2020 contains four domains (herbarium, herbarium_photo_associations, photo and test). The herbarium domain contains 320,750 images in 997 species, and the number of images in different species are unbalanced. This dataset consists of herbarium sheets whereas the test set will be composed of field pictures. The validation set consists of two domains herbarium_photo_associations and photos. Herbarium_photo_associations domain includes 1,816 images from 244 species. This domain contains both herbarium sheets and field pictures for a subset of species, which enables learning a mapping between the herbarium sheets domain and the field pictures domain. Another photo domain has 4,482 images from 375 species and images are from plant pictures in the field, which is similar to the test dataset. The test dataset contains 3,186 unlabeled images.


Results

[Paper]

[Bibtex]