This code is for the work "Complex Dual-Tree Pyramid Scattering Transformer"
For detailed environment configuration requirements, please look at the requirement.txt file.
Detailed data preprocessing, hyperparameter selection, and training Settings are as follows：
- Image Classification Task
CIFAR Dataset
Data Preprocessing: CIFAR-10 is a small image classification data set containing 10 categories，each category consists of 50,000 training and 10,000 testing images. The color image with a resolution of $32×32$. CIFAR-100 is a classification data set with 100 categories but fewer samples (500 for training and 100 for testing) with the same resolution. The input with the resolution of $32×32$ in the initial pyramid branch is padded to $48×48$, and then divided into parallel inputs of size $[48×48, 24×24, 12×12]$.
Hyperparameter Settings: Parallel scattering input branches index $b$ is set to $[2,3,4]$, the scale scattering parameter $q$ is configured as $[1,2,2]$, the scattering dimensions are set to $[21,129,129]$, and the scattering learning rate (lr_scattering) is set to $0.1$.
Training: Training starts from scratch with a learning rate of $0.1$, a maximum learning rate of $6e^{-2}$, and 300 total epochs. Weight decay and momentum are set to $5e^{-4}$ and $0.9$, respectively.
Tiny-ImageNet Dataset
Data Preprocessing: Tiny-Imagenet consists of 100k training images, 10k validation images, and 10k test images belong to 200 categories with a resolution of $64×64$ pixels. The $64×64$ input in the initial pyramid branch is padded to $80×80$ and then divided into parallel inputs of size $[80×80, 40×40, 20×20]$.
Hyperparameter Settings: Parallel scattering input branches index $b$ is set to $[3,4,5]$, the global scale scattering parameter $q$ is set to $2$, the scattering dimensions are uniformly set to $[129,129,129]$, and lr_scattering is $0.1$.
Training: Network is randomly initialized and trained with a cosine decay learning rate scheduler. The initial learning rate is $5e^{-4}$, the weight decay is $5e^{-2}$, and the warm-up decay is $10$ epochs. The total epochs are set to 300.
COVIDx CRX-3 Dataset
Data Preprocessing: COVIDx CXR-3 is a medical chest X-ray detection data set containing chest X-ray images of COVID-19 patients from a multinational cohort with a resolution of $1024×1024$. There are 29986 training set images and 400 testing set images. Input images in the initial branch are reshaped to $384×384$ and divided into parallel inputs of sizes $[96×96, 48×48, 24×24]$ with a pyramid sequence of patch sizes $(16, 8, 4)$.
Hyperparameter Settings: We take more levels of scattering components for aggregation, and $q=3$ is applied to branches $b=[4,5]$, while $q=2$ is retained for $b=3$.
Training: The networks are trained for $400$ epochs using a maximum learning rate of $1e^{-2}$. The learning rate scheduler is set to OneCycleLR. The weight decay is set to $5e^{-4}$. The momentum is set to $0.9$.
- Object Tracking Task
TrackingNet Dataset
Data Preprocessing: TrackingNet is a large-scale tracking dataset including 30k videos withmore than 14 million 780 annotations. We evaluate our CPST-Tracker on 511 official test sequences. To maintain spatial consistency of scattering features, we pad the template input to the same size as the search region $224×224$.
Hyperparameter Settings: The number of hidden channels per stage is configured as $[48, 192, 192, 384]$, corresponding layer counts set to $[2, 2, 4, 2]$.
Training: The CPST-Tracker is trained on LaSOT, TrackingNet, GOT-10k, and COCO 2017 training splits. Optimization is performed using AdamW with a learning rate of $5e^{-4}$, weight decay of $1e^{-4}$, and a backbone-specific learning rate of $5e^{-5}$. The total epochs are set to 300. DropPath is applied with a rate of $0.1$.