Object detection using R-CNN

Deep Learning for Images with PyTorch

Michal Oleszak

Machine Learning Engineer

Region-based CNN family: R-CNN

R-CNN family: R-CNN, Fast-CNN, Faster CNN

R-CNN

Module 1: generation of region proposals

¹ Citation: Jason Brownlee. 2019. Deep Learning for Computer Vision.

Region-based CNN family: R-CNN

R-CNN family: R-CNN, Fast-CNN, Faster CNN

R-CNN

Module 1: generation of region proposals
Module 2: feature extraction (convolutional layers)

¹ Citation: Jason Brownlee. 2019. Deep Learning for Computer Vision.

Region-based CNN family: R-CNN

R-CNN family: R-CNN, Fast-CNN, Faster CNN

R-CNN

Module 1: generation of region proposals
Module 2: feature extraction (convolutional layers)
Module 3: class and bounding box prediction

¹ Citation: Jason Brownlee. 2019. Deep Learning for Computer Vision.

R-CNN: backbone

Convolutional layers: pre-trained models
- Backbone: the core CNN architecture responsible for feature extraction

backbone

Convolutional & pooling layers
Extract features for region proposals and object detection

R-CNN: backbone with PyTorch

import torch.nn as nn
from torchvision.models import vgg16,
    VGG16_Weights


vgg = vgg16(weights=VGG16_Weights.DEFAULT)

vgg model

R-CNN: backbone with PyTorch

import torch.nn as nn
from torchvision.models import vgg16,
    VGG16_Weights


vgg = vgg16(weights=VGG16_Weights.DEFAULT)

vgg model features

.features: only convolutional layers

R-CNN: backbone with PyTorch

import torch.nn as nn
from torchvision.models import vgg16,
    VGG16_Weights


vgg = vgg16(weights=VGG16_Weights.DEFAULT)

vgg model

.features: only convolutional layers
.children(): all layers from block

R-CNN: backbone with PyTorch

import torch.nn as nn
from torchvision.models import vgg16,
    VGG16_Weights


vgg = vgg16(weights=VGG16_Weights.DEFAULT)


backbone = nn.Sequential(
    *list(vgg.features.children())
)

nn.Sequential(*list()): all sub-layers are placed into a sequential block as a list
- *: unpacks the elements from the list

vgg model

.features: only convolutional layers
.children(): all layers from block

R-CNN: classifier layer

Extract backbone's output size

input_dimension = nn.Sequential(*list(
    vgg_backbone.classifier.children())
)[0].in_features

Create a new classifier

classifier = nn.Sequential(
    nn.Linear(input_dimension, 512),
    nn.ReLU(),
    nn.Linear(512, num_classes),
)

R-CNN: box regressor layer

Sits on top of the backbone
4 outputs for the 4 box coordinates

box_regressor = nn.Sequential(
    nn.Linear(input_dimension, 32),
    nn.ReLU(),
    nn.Linear(32, 4),
)

Putting it all together: object detection model

class ObjectDetectorCNN(nn.Module):
    def __init__(self):
        super(ObjectDetectorCNN, self).__init__()

        vgg = vgg16(weights=VGG16_Weights.DEFAULT)
        self.backbone = nn.Sequential(*list(vgg.features.children()))

        input_features = nn.Sequential(*list(vgg.classifier.children()))[0].in_features

        self.classifier = nn.Sequential(
            nn.Linear(input_features, 512),
            nn.ReLU(),
            nn.Linear(512, 2),
        )

        self.box_regressor = nn.Sequential(
            nn.Linear(input_features, 32),
            nn.ReLU(),
            nn.Linear(32, 4),
        )

Putting it all together: object detection model

class ObjectDetector(nn.Module):
    (...)

    def forward(self, x):

        features = self.backbone(x)

        bboxes = self.regressor(features)
        classes = self.classifier(features)
        return bboxes, classes

Running object recognition

Load and transform the image
unsqueeze() the image to add the batch dimension
Pass the image tensor to the model
Run Non-Max Suppression (nms()) over model's output
draw_bounding_boxes() on top of the image

Let's practice!

Deep Learning for Images with PyTorch