flutterby / README.md
flobbit's picture
Update README.md
fdccaec
|
raw
history blame
1.79 kB
metadata
license: apache-2.0
tags:
  - en
  - image classification
  - fastai
model-index:
  - name: flutterby by flobbit
    results:
      - task:
          name: image classification
          type: image-classification
        metrics:
          - name: accuracy
            type: acc
            num_train_epochs: 10
            learning_rate: 0.00363
            value: 77.3
metrics:
  - accuracy
pipeline_tag: image-classification

FlutterBy ST Swallowtail Butterfly Insect Classification

Model description

The model is used to classify images into one of the 51 North American swallowtail or cattleheart butterfly species. resnet50 was used for training.

Intended uses & limitations

The model was trained on 8577 insect images spread over 51 species. The model is likely biased toward some species being more commonly found in certain habitats.

Training and evaluation data

The images used in training were obtained from GBIF: GBIF.org (22 June 2023) GBIF Occurrence Download https://doi.org/10.15468/dl.bqg8bw

Only the first 400 images of each species (if available) were downloaded. The image set was partially cleaned for quality to remove caterpillars, poor images or butterflies that were too far away for proper ID. After "cleaning", 200 additional images were downloaded for Battus philenor and Battus polydamas (as those species had a very high percentage of caterpillar shots).

The dataset is primarily "in the wild" shots rather than all staged poses, and includes images for which even an expert would not be able to see identifying characteristics (hence the lower overall accuracy).

The image set had a minimum of 30 pics in a class for the less uncommon species (which is not enough for accurate training but they were included for completeness). 33 species had over 200 images (after cleaning).