Vision transformer pre-trained with MAE on Formula 1 racing dataset

Vision transformer baze-sized (ViT base) feature model. Pre-trained with Masked Autoencoder (MAE) Self-Supervised approach on the custom Formula 1 racing dataset from Constructor SportsTech, allowes the extraction of features that are more efficient for use in Computer Vision tasks in the areas of racing and Formula 1 then features pre-trained on standard ImageNet-1K. This ViT model is ready for use in Transformers library realization of MAE.

Model Details

Model type: feature backbone
Image size: 224 x 224
Original MAE repo: https://github.com/facebookresearch/mae
Original paper: Masked Autoencoders Are Scalable Vision Learners (https://arxiv.org/abs/2111.06377)

Training Procedure

F1 ViT-base MAE was pre-trained on the custom dataset containing more than 1 million Formula 1 images from seasons 2021, 2022, 2023 with both racing and non racing scenes. The traing was performed on a cluster of 8 A100 80GB GPUs provided by Nebius who invited us to technical preview of their platform.

Training Hyperparameters

Masking proportion during pre-training: 75 %
Normalized pixels during pre-training: False
Epochs: 500
Batch size: 4096
Learning rate: 3e-3
Warmup: 40 epochs
Optimizer: AdamW

Comparison with ViT-base MAE pre-trained on ImageNet-1K

Comparison of F1 ViT-base MAE and original ViT-base MAE pre-trained on ImageNet-1K by reconstruction results on images from Formula 1 domain. Top is F1 ViT-base MAE reconstruction output, bottom is original ViT-base MAE.

How to use

Usage is the same as in Transformers library realization of MAE.

from transformers import AutoImageProcessor, ViTMAEForPreTraining
from PIL import Image
import requests

url = 'https://huggingface.co/andrewbo29/vit-mae-base-formula1/blob/main/racing_example.jpg'
image = Image.open(requests.get(url, stream=True).raw)

processor = AutoImageProcessor.from_pretrained('andrewbo29/vit-mae-base-formula1')
model = ViTMAEForPreTraining.from_pretrained('andrewbo29/vit-mae-base-formula1')

inputs = processor(images=image, return_tensors="pt")
outputs = model(**inputs)
loss = outputs.loss
mask = outputs.mask
ids_restore = outputs.ids_restore

BibTeX entry and citation info

@article{DBLP:journals/corr/abs-2111-06377,
  author    = {Kaiming He and
               Xinlei Chen and
               Saining Xie and
               Yanghao Li and
               Piotr Doll{\'{a}}r and
               Ross B. Girshick},
  title     = {Masked Autoencoders Are Scalable Vision Learners},
  journal   = {CoRR},
  volume    = {abs/2111.06377},
  year      = {2021},
  url       = {https://arxiv.org/abs/2111.06377},
  eprinttype = {arXiv},
  eprint    = {2111.06377},
  timestamp = {Tue, 16 Nov 2021 12:12:31 +0100},
  biburl    = {https://dblp.org/rec/journals/corr/abs-2111-06377.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}

andrewbo29
/

vit-mae-base-formula1