Spatial Pyramid Pooling helps the network output the same shape regardless of any aspect ratio and input size. Instead of Pooling with a fixed filter size, it divides the input with different levels of ratio, so the output would not…
Unet may be one of the most basic researches on segmentation tasks. It consists of 3 parts: Encoding Phase (Apply Convolutions to classify object) -> Bridge -> Decoding Phase(Restore information so that the output would be 388×388). During the final…
HRNet was a research done by Microsoft which lead to higher performance compared with state-of-the-art architectures. Traditional segmentation models utilize skip connections in order to recover spatial information from previous layers. The problem with this method is that it can’t…
In order to classify images more precisely, the traditional way is to apply convolution and pooling to lower the dimension of the input so that the model can understand more complex features. This is ok for classification tasks because you…
Attention Unet highlights only relevant activations during training. This can not only perform better when the target you want to detect is relatively tiny compared to the size of the picture, but it can also reduce unnecessary computations. The overall…
If I were to restart a semantic segmentation model project again, I would choose the loss function by considering the following. Is there any imbalance in your data? ⇒（NO）⇒ Binary Cross Entropy ⇓ (YES) ⇓ Is the area you want…
I did an experiment to compare how model performance change when converted to TensorRT engine. These are the steps I took. 1. Train Several Binary Segmentation Model 2. Randomly Choose 30 images from the dataset 3. Run inference with the…
Argmax compares pixels in the same position across channels, and acquires the index of the highest channel. This can be useful for semantic segmentation. Semantic segmentation models outputs the same width and height as the input image and creates a…
I found out why my TensorRT engine model was not working as expected. I messed up with configuring the preprocessing step for Deepstream. When you use Deepstream to run inference, there is a property called net-scale-factor and offsets which you…
I’ve compared the performance between the model before and after converting to TensorRT engine. The output from the TensorRT engine is flattened to a 1d array, so I had to convert that back to a 2d array to display the…