WildMAE extends masked autoencoders for wildlife camera-trap images
Original title: WildMAE: Structure-Aware Masked Autoencoding for Wildlife Image Reconstruction
On 4,744 Snapshot Serengeti images at 75% masking, historical PSNR was 27.90 dB for WildMAE versus 24.33 for vanilla MAE.
WildMAE is a decoder-side masked-autoencoder extension for camera-trap imagery, combining local structure-aware refinement with masked-to-visible cross-attention. A zero-gated convolutional residual adds a two-dimensional neighborhood prior, while gated cross-attention retrieves context from visible tokens without replacing the Transformer decoder. Training progressively adds normalized-pixel, Sobel-edge, and perceptual objectives. The authors treat the cited Snapshot Serengeti scores as historical and protocol-dependent, not a new harmonized comparison.
Unconfirmed: Cited scores are historical protocol-dependent results, not a newly reproduced comparison or a test of restoring unavailable pixels.
- Historical SSIM was 0.840 for WildMAE versus 0.780 for vanilla MAE.
- Historical LPIPS was 0.160 versus 0.320.
- The recipe adds normalized-pixel, Sobel-edge, and perceptual objectives.
- Authors also cite favorable averages under four synthetic degradations.
- AI models
- WildMAE
From the original article
1 min read at arxiv.orgCamera-trap imagery presents substantial variation in animal appearance, illumination, background clutter, and visual degradation. Masked autoencoders (MAEs) offer a label-free framework for modeling such data through reconstruction of randomly hidden image patches. We present WildMAE, a decoder-side extension of MAE that combines local structure-aware refinement with masked-to-visible cross-attention.
By Yicheng Zhang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.