ScaleNet – Improve CNNs through Recursively Rescaling Objects
Authors: Xingyi Li, Zhongang Qi, Xiaoli Fern, Fuxin Li11426-11433
AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To demonstrate the robustness of our proposed framework, we conduct experiments with pretrained as well as fine-tuned classification and detection frameworks on MNIST, CIFAR-10, and MS COCO datasets and results reveal that our proposed framework significantly boosts the performances of deep networks. |
| Researcher Affiliation | Collaboration | Xingyi Li,1 Zhongang Qi,1,2 Xiaoli Z. Fern,1 Li Fuxin1 1School of Electrical Engineering and Computer Science, Oregon State University 2Applied Research Center, PCG, Tencent |
| Pseudocode | Yes | Algorithm 1 Recursive-Scale Net |
| Open Source Code | No | The paper does not provide an explicit statement or link to the source code for the methodology described in the paper. It only references a third-party benchmark (maskrcnn-benchmark) that they used. |
| Open Datasets | Yes | MNIST contains 60, 000 handwritten digits in the training set and 10, 000 digits in the test set, coming from 10 distinct classes. The original size of each image is 28 28. ... In CIFAR-10, there are 60, 000 RGB images in 10 classes in the training set, and 10, 000 images in the test set. The original size of each image is 32 32. ... In COCO, there are about 118K training images containing objects from 80 classes. The validation set contains 5K images, and the test-dev set contains about 22K. |
| Dataset Splits | Yes | The same rescaling operations are performed on MNIST test set to generate the validation set. ... The test set is built through rescaling original validation images to the size of 96 96. ... The validation set contains 5K images, and the test-dev set contains about 22K. |
| Hardware Specification | Yes | FPS (frames per second) is reported based on a single 2080 Ti GPU. |
| Software Dependencies | No | The paper mentions software components and tools like U-Net, VGG-16, Adam optimizer, and Focal loss, and refers to Facebook Detectron, but does not provide specific version numbers for these software dependencies (e.g., PyTorch 1.9, TensorFlow 2.x, or specific library versions). |
| Experiment Setup | Yes | The batch size is 64 and the optimizer is Stochastic Gradient Descent (SGD) with momentum with learning rate 0.01. ... The RH in algorithm 1 is set to 1.05, and k is set to 15%. ... Training images are resized to 512 512 before feeding to Scale Net. ... Focal loss (Lin et al. 2017b) is applied for class imbalance with the parameter γ = 2. The optimizer is Adam (Kingma and Ba 2014) with 1e 4 learning rate, and the batch size is 16. |