I recently implemented MagicPoint, the base detector from the paper SuperPoint: Self-Supervised Interest Point Detection and Description, and was excited to achieve an mAP of 0.9721, which closely matches the original paper’s reported 0.971, validating the effectiveness of my implementation.
This journey involved overcoming various technical challenges, optimizing training, and uncovering key insights into deep learning practices. Below, I share my results and important lessons learned along the way.
Here’s a visualization of MagicPoint’s detections on several synthetic images:
Below are the mean average precision, mean localization error, and validation loss curves from training:
Mean Average Precision
Mean Localization Error
Validation Loss
By default, TensorFlow initializes CNN layers using Glorot Uniform, but for ReLU activations, the He Normal initializer is more effective. Switching to He Normal improved training stability and convergence.
I used to normalize images by dividing pixel values by 255, assuming it was sufficient. However, after reading Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification, I realized that proper variance propagation across layers requires inputs to be standardized (mean = 0, std = 1). This significantly impacts training stability and performance.
- SuperPoint: Self-Supervised Interest Point Detection and Description
- Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification
This project was a great learning experience! From achieving state-of-the-art results to understanding initialization and input normalization, this journey has deepened my knowledge of deep learning best practices.
Feel free to check out the code, experiment with it, and contribute! 🚀
📩 Questions or feedback? Open an issue or start a discussion!