Rectified Activations for Regression Problems

This repository contains example code for the poster presentation "ReDepth" at The 12th International Conference on Robot Intelligence Technology and Applications (RiTA 2024).

This research focuses on the abnormal activation issue, where the skewness of activations increases when estimating out-of-domain (OOD) data compared to the originally trained domain. The study demonstrates that performance can be improved by addressing this issue through a clipping method for activations.

The example codes provided include various regression problems, illustrating that the proposed ReAct (Rectified Activations) method enhances estimation performance in OOD scenarios.

---

Examples Included

1. Sine Wave Example

This code trains a neural network on a sine wave and evaluates its performance when input values fall outside the trained range.

2. MNIST Example

MNIST is a well-known handwritten digit dataset.
In this example:

• The model is trained on the original MNIST dataset.
• OOD data is generated by applying transformations such as rotation, translation, and noise.
• The ReAct method demonstrates superior results when testing with OOD data.

3. Depth Estimation Example

This example focuses on depth estimation using the KITTI dataset for training. The trained network is then tested on DIODE and Make3D datasets.

• mono_diode.py and mono_make3d.py: Results without applying ReAct.
• mono_react_diode.py and mono_react_make3d.py: Results with ReAct applied.

Data Acquisition:

• DIODE Dataset: Automatically downloaded during code execution.
• Make3D Dataset: Can be downloaded using the Make3D dataset link, which provides 400 training images and 300 aligned depth maps.