This repository contains the implementation method of our Herbarium-Field Triplet Loss Network (HFTL Network) and One-streamed Mixed Network (OSM Network) in the context of Cross-Domain Plant Identification. Our results show that the HFTL Network can generalize rare species as equally as species with many training data better than the OSM Network (conventional CNNs).
Figure A and B below show the Top-5 predictions of a plant sample with its predicted scores and activation maps from the HFTL and OSM Networks respectively. More samples of comparison can be found here.
HFTL Network | OSM Network |
---|---|
A machine learning approach for cross-domain plant identification using herbarium specimens
https://doi.org/10.1007/s00521-022-07951-6
- Tensorflow 1.12
- TensorFlow-Slim library
- imagenet pretrained models (inception v4 and inception resnet v2)
- HFTL Network
- OSM Network
-
HFTL Network
-
OSM Network
- HFTL Network
- OSM Network
- clef2021_shared_class_herbarium_train.txt
- clef2021_shared_class_herbarium_test.txt
- clef2021_shared_class_field_train.txt
- clef2021_shared_class_field_test.txt
- test_set_1_seen.txt (with field training images)
- test_set_2_unseen.txt (without field training images)
(note: the image files from Test Set 2 are sourced from Google Image queries)
- HFTL Network
- OSM Network