본문 바로가기

fatal factors that make the self-supervised training for monodepth estimation extremely difficult when trained on nuScenes Human: I've been working on monodepth estimation in a self-supervised manner and have referred to monodepth2. While training on nuScene, I found it hard to train the model for some reason. What would be fatal factors for the sucessful training? ChatGPT: Training a self-supervised monodepth model like Monodepth2 on a complex dataset like nuScenes can be tricky. Here are some fatal factors that co.. 더보기
Noise estimation idea 2 Human: Assume we want to OGMs for 2 classes, which results in {X_1, X_2}. When estimating the intrinsic noise in X1, I think I can take advantage of X_2 as a prior information. What would be a common practice? ChatGPT: Leveraging X2X_2X2​ as Prior Information for Noise Estimation in X1X_1X1​Since you are estimating the intrinsic noise in X1X_1X1​, it makes sense to use X2X_2X2​ as prior informat.. 더보기
Improving DDPD noise estimation ChatGPT: To improve the performance of your U-Net-based intrinsic noise estimation network, consider the following architectural modifications: 1. Enhancing Feature Representationsa. Using Residual Connections (ResUNet)Replace standard convolutional blocks with Residual Blocks to ease gradient flow and improve training stability.Residual blocks reduce vanishing gradient issues, helping the netwo.. 더보기