VideoMAE Large Finetuned Kinetics
| Property | Value |
|---|---|
| Parameter Count | 304M |
| License | CC-BY-NC-4.0 |
| Paper | VideoMAE Paper |
| Accuracy | 84.7% (Top-1), 96.5% (Top-5) |
| Framework | PyTorch |
What is videomae-large-finetuned-kinetics?
VideoMAE is a sophisticated video classification model that extends the Masked Autoencoder (MAE) approach to video understanding. This large variant has been pre-trained for 1600 epochs using self-supervised learning and then fine-tuned on the Kinetics-400 dataset. The model processes videos as sequences of 16x16 fixed-size patches and employs a Vision Transformer architecture with additional decoder capabilities.
Implementation Details
The model architecture builds upon the Vision Transformer (ViT) framework, incorporating several key technical innovations:
- Uses a [CLS] token for classification tasks
- Employs sinus/cosinus position embeddings
- Processes video inputs as 16x16 patch sequences
- Implements a transformer encoder-decoder architecture
- Utilizes masked autoencoding for self-supervised learning
Core Capabilities
- Video classification across 400 Kinetics categories
- Feature extraction for downstream tasks
- High-accuracy prediction (84.7% top-1)
- Efficient processing of video sequences
- Robust representation learning
Frequently Asked Questions
Q: What makes this model unique?
This model stands out for its application of masked autoencoding to video data, achieving state-of-the-art performance while using an efficient self-supervised pre-training approach. The large parameter count (304M) and impressive accuracy metrics make it particularly suitable for complex video understanding tasks.
Q: What are the recommended use cases?
The model is primarily designed for video classification tasks, particularly within the Kinetics-400 dataset categories. It's well-suited for applications requiring high-accuracy video understanding, such as content categorization, action recognition, and video indexing systems.
