The document discusses object detection pipelines. It begins by defining object detection as identifying objects in images and locating them with bounding boxes. The main components of an object detection pipeline are datasets, preprocessing, model selection and training, testing and evaluation. Popular models discussed are Faster R-CNN, R-FCN, and SSD which use deep convolutional neural networks as feature extractors and classifiers. Key evaluation metrics are mean average precision and prediction time/memory usage. Popular datasets mentioned are MSCOCO, Pascal VOC, and LSVRC. The document provides information on preprocessing, training including fine-tuning pre-trained models, and codes/models available on GitHub.

1. Object Detection Pipeline
Abhinav Dadhich, ABEJA, Inc.
Tokyo Machine Learning Kitchen, 2017

2. Outline
1. What is Object Detection? Components?
2. Models for detection
3. Train, Test & Evaluate
4. FAQs

3. Object Detection
http://cs231n.stanford.edu/slides/winter1516_lecture8.pdf
Q: What object is in image?
A: Cat

4. Object Detection
http://cs231n.stanford.edu/slides/winter1516_lecture8.pdf
Q: Where the object is in image?
A: bounding box or coordinates

5. Object Detection
http://cs231n.stanford.edu/slides/winter1516_lecture8.pdf

6. Components of
a Detection
● Dataset of images and
target labels
● Pre-Processing images
and labels
● Model selection and
modifications
● Training
● Testing and Evaluation
● Deploying final model

7. Model Architecture
Feature
Extractor
Classifier
Object
Classification
Bounding Box

8. Why Deep Neural Network?
“Fast R-CNN” Ross Girshick, ICCV’15

9. Model Architecture
Feature
Extractor
Classifier
Object
Classification
Bounding Box
Deep CNN Architecture
● VGG
● Inception
● Resnet

10. Faster Region based ConvNets (Faster-RCNN)
Girshick et al. ICCV 2015.
Fig: Huang et al. 2016, arXiv:1611.10012v1
● 2 Step Process
● Higher Accuracy
● Slower per
sample prediction
wrt similar
models.
● Large image size

11. Region based Fully Convolutional Network(R-FCN)
Dai et al. , NIPS 2016.
● 2 Step Process
● Faster per sample
prediction time wrt
Faster-RCNN
Fig: Huang et al. 2016, arXiv:1611.10012v1

12. Single Shot Detector(SSD)
Liu et al. , ECCV 2016
● 1 step Process
● Faster per sample
prediction time
● Small images,
large objects
Fig: Huang et al. 2016, arXiv:1611.10012v1

13. Training
● Pre-trained models are available for major deep learning
frameworks.
● Fine tune existing model
Feature Layers
CNNs FCs
Classification Layers

14. Training
● Pre-trained models are available for major deep learning
frameworks.
● Fine tune existing model, re-initializing output layers.
Feature Layers
CNNs FCs
Re-initialized
Classification Layers

15. Evaluation Metrics & Tests
● Mean Average Precision(mAP):
○ Thresholding based on Intersection over Union(IoU) score.
○ Average over all class predictions.
○ Higher is better.
● Prediction Time : pre-processing + prediction time per image
● Memory Usage : model’s gpu/cpu memory usage while
prediction

16. Datasets
● [Try] Collect new dataset according to task.
● [Try] Get labels as accurate as possible.
● If not, use public datasets:
○ MSCOCO: 80 objects, 300K images, 5 captions per image.
○ Pascal VOC: 20 objects, ~20K images
○ LSVRC: 200 objects, ~470K images

17. Pre-processing
● Image resizing.
● Image pixel normalization.
● Data Augmentation:
○ Flip
○ Random rotations

18. Codes and Pre-trained Models
● Faster-RCNN: https://github.com/rbgirshick/py-faster-rcnn (Caffe)
● R-FCN : https://github.com/Orpine/py-R-FCN (Caffe)
● SSD: https://github.com/weiliu89/caffe/tree/ssd (Caffe)

19. Thank You