- The Hidden Layer
- Posts
- Learning Computer Vision Week 6
Learning Computer Vision Week 6
Week 6 of documenting my AI/ML learning journey (Oct 13 - Oct 19)
Brandon Kim
November 16, 2024
What was discussed last week…
Strategies to how CV models conduct object detection
The reliability of Haar-Cascade Classifiers
The trend of how almost any difficult math concept can be used in CS “under the hood”
This week’s newsletter is a short one, sorry guys. 😓
Tuesday, October 15
For the last part of my Computer Vision learning, I experimented with Faster R-CNNs, in which the “R” stands for “region-based”. This type of CNN reduces the running time of object detection networks by combining RPNs (region proposal networks) and CNNs, along with other components; GeeksForGeeks provides a great explanation for how Faster R-CNNs work, and here is the actual proposal for the new model.
For making a Faster R-CNN, the IBM lab I was in used torchvision, a library that is a part of PyTorch.
model_ = torchvision.models.detection.fasterrcnn_resnet50_fpn(pretrained=True)This function pretrained on the COCO dataset, has a few parameters but the most notable ones are:
pretrained bool (as shown): where you can decide to use the pretrained version of torchvision’s Faster R-CNN (which uses COCO).
weights FasterRCNN_ResNet_FPN_Weights: the function uses no pre-trained weights by default, but torchvision includes a FasterRCNN_ResNet_FPN_Weights parameter value that you can just insert in there if you wanted to have a set of pre-trained weights.
num_classes int: The number of output classes of the model.
Lessons Learned
I learned about the COCO dataset and learned about the .fasterrcnn_resnet50_fpn() function that uses it
Resources
Course I followed:
