Deep Learning - Feature Pyramid Network - How to understand the downsampling notation?
I have a question concerning the notation of the downsampling process in the feature pyramid network (FPN) architecture. I'm not sure, whether stack overflow is actually the best place for this question. Any hints concerning better places are therefore very welcome.
My question can best be illustrated with the following image from a presentation of one of the original authors of FPN:
Source: http://presentations.cocodataset.org/COCO17-Stuff-FAIR.pdf, Slide 11
The scale annotations of 1 and 1/4 make sense to me. Obviously, we start at full scale and after one pooling step, we have a scale of 1/4, because we downsized by a factor of 2 in the x- and y- directions. But as far as I understand, following the same logic, at the next stage (i.e. after the next pooling), we should have a scale of 1/16. After the next step 1/64, etc. What am I missing?
After one polling step you will get a scale of 1/2 and not 1/4. The scale reffers to the change along an axis and not the ratio of the areas. So why you have a change of 1/4 in the beginning? As slide 11 states the drawing reffers to resnet/resnext model. If we look at the resnet model architecture we can see that first we have a convolution with 7x7 with stride 2 and following this we have polling layer with stride 2 so over all we get a 1/4 reduction per axis. In the next stages we only have the polling with stride 2 so we only get a change of factor 2. I.e 1/8, 1/16 1/32.
- tensorflow:Can save best model only with val_acc available, skipping
- Issue with training Keras model using ModelCheckpoint in Kaggle notebook (Unexpected result of `train_function` (Empty logs))
- How to fix the learning-rate for Huggingface´s Trainer?
- What is the difference between backpropagation and reverse-mode autodiff?
- why softmax get small gradient when the value is large in paper 'Attention is all you need'
- What does model.eval() do in pytorch?
- PyTorch RuntimeError: device >= 0 && device < num_gpus INTERNAL ASSERT FAILED
- How to add the bounding box values to the labels text files during prediction with a trained YOLO-V8 instance segmentation model?
- How to calculate optimal batch size?
- What is FLOPS in field of deep learning?
- Tensorflow TFLearn working example with TFLearn
- Converting Detectron2 to other frameworks
- How can I use a pre-trained neural network with grayscale images?
- How to deep-learn simple data in exponential form?
- Is GradScaler necessary with Mixed precision training with pytorch?
- How to avoid "CUDA out of memory" in PyTorch
- PyTorch model input shape
- In MLflow, is there a way group metrics & params into "sections" like the system section?
- Difference between TensorFlow model fit and train_on_batch
- SNPE Model conversion: Syntax for model input with 0 dimensions
- Loss remains unchanged in a multiclass classification deep model
- Python tensorflow keras error when load a json model: Could not locate class 'Sequential'
- What is validation set for yolov5 training and is it necessary?
- stmemory and ltmemory in "How to build your own AlphaZero AI using Python and Keras"
- YOLOv8: How to calculate map on test set
- In pytorch, how to use the weight parameter in F.cross_entropy()?
- Low accuracy i.e. 52% in predicting whereas the training and validation accuracy during training is around 92%
- Concatenate two layers in keras, tensorflow
- Implementing dropout from scratch
- UNET prediction