What does "TDSM" mean?
Table of Contents
TDSM is a method used for recognizing actions based on skeleton data while dealing with situations where the actions have not been seen before. The main idea behind TDSM is to match the features of skeleton data with text descriptions of actions. This helps the system predict new actions accurately.
Traditional methods have had trouble because the skeleton data and text data don't easily fit together, which makes it challenging to learn effectively. TDSM takes inspiration from how some text-to-image models work and focuses on improving the training process. Instead of generating new data, it aligns the skeleton features with the text descriptions during training.
To make the matching process stronger, TDSM uses a special technique called triplet diffusion loss. This helps ensure the correct matches between skeleton and text are closer together, while the incorrect ones are pushed further apart. As a result, TDSM shows better performance than current top methods, allowing it to make more accurate predictions in situations where it encounters new actions.
Label-Noise Robust Diffusion Models
Training models to generate data often requires large datasets, which can include noisy or incorrect information. This noise can harm the quality of the generated data and make it less reliable.
To address this issue, a new approach called TDSM is introduced. This method focuses on training these models in a way that manages the noise present in the dataset. It combines different scoring methods and considers how labels may change over time. By customizing the process to account for these noise factors, TDSM improves the quality of the generated outputs.
Through testing across different datasets, TDSM has shown that it can produce better samples that align more closely with the intended conditions. This method also demonstrates improved performance over standard ways of correcting noisy labels, proving to be an effective strategy for training models in the presence of noise.