Bird Species Classification

Victor Chan · 104384986 · COS30082 Assignment 1

Upload a bird photo and a CNN names its species, out of 200 (CUB-200). The final model is ConvNeXt-Tiny. Every other model trained along the way can be picked too, down to the baselines. All of them run in your browser, so the photo never leaves your device.

or drop or paste an image anywhere on the page.

Prediction

Loading the model

Starting the download

    Example photo

    How it works

    1. Photo Short side resized to 512 px, middle 448 × 448 kept
    2. ConvNeXt-Tiny A CNN with 28 million weights: layers of filters find edges, then textures, then bird parts
    3. 200 probabilities Softmax turns its 200 scores into probabilities that sum to 100%
    4. Answer The species with the highest probability

    Where its knowledge comes from. ConvNeXt-Tiny first learned general shapes and textures from 1.28 million ImageNet photos. It was then fine-tuned on the 4,829 bird training photos for 30 epochs: every layer learned, the copied layers at a learning rate 10 times smaller than the new last layer, so they kept most of what they knew.

    How it avoids memorising. With about 24 photos per species, a network can memorise the training photos instead of learning the birds. Four tools work against that, together: strong augmentation (random crops, turns and colour changes), label smoothing, weight decay and dropout.

    Why ConvNeXt-Tiny.

    Key decisions

    Training

    Results

    Model Test Top-1 Average per-class

    Limitations

    • It knows only 200 species. Any other bird, or a photo with no bird, still gets one of the 200 names.
    • It learned from about 24 photos per species, and it sees only the middle square of the photo.