Posts
-
Detection and Drivable road segmentation
This post showcases the training result of a combination of SSD and FC8 network. Modified SSD which classifies 41 classes and FC8 network that outputs 3 types of road.
-
文本检测与识别 CTPN && CRNN && STN
总结
- 端到端文本检测，端到端日期识别
- 不需要进行字符分割和水平缩放，可识别任意长度序列
- CNN+RNN架构，检测结果更具鲁棒性
- CTC选择最优序列
- 可检测部分旋转后的日期
- 无需对旋转后的角度进行数据标注
- STN + CTPN + CRNN
-
YOLO vs Faster RCNN
This post talks about YOLO and Faster-RCNN. These are the two popular approaches for doing object detection that are anchor based. Faster RCNN offers a regional of interest region for doing convolution while YOLO does detection and classification at the same time. I would say that YOLO appears to be a cleaner way of doing object detection since it’s fully end-to-end training. The Faster RCNN offers end-to-end training as well, but the steps are much more involved. Nevertheless I will describe both approaches carefully in detail.
-
A comparison between VAE and GAN
This post concludes VAE and GAN
I’ve took some time going over multiple post regarding VAE and GAN. To help myself to better understand these generative model, I decided to write a post about them, comparing them side by side. Also I want to include the necessary implementation details regarding these two models. For this model, I will use the toy dataset which is MNIST. The code in this post will be mainly implemented with Tensorflow.