[go: up one dir, main page]
Scribd
Upload
75 views24 pages

Introduction To Object Detection

This document introduces object detection in computer vision. It discusses classification and localization tasks, and how object detection combines these by outputting class labels and bounding boxes. Evaluation metrics for object detection include intersection over union (IoU) for localization accuracy and mean average precision (mAP) across classes based on the precision-recall curve. Popular datasets that have driven progress are also presented, from Pascal VOC to ImageNet to MS COCO. State-of-the-art models and evaluation metrics continue to evolve for robust object detection.

Uploaded by

i not
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
75 views24 pages

Introduction To Object Detection

This document introduces object detection in computer vision. It discusses classification and localization tasks, and how object detection combines these by outputting class labels and bounding boxes. Evaluation metrics for object detection include intersection over union (IoU) for localization accuracy and mean average precision (mAP) across classes based on the precision-recall curve. Popular datasets that have driven progress are also presented, from Pascal VOC to ImageNet to MS COCO. State-of-the-art models and evaluation metrics continue to evolve for robust object detection.

Uploaded by

i not
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 24

Introduction to Object Detection

Computer Vision Tasks

Source: CS231n Object detection http://cs231n.stanford.edu/slides/2016/winter1516_lecture8.pdf


Computer Vision Tasks

Source: CS231n Object detection http://cs231n.stanford.edu/slides/2016/winter1516_lecture8.pdf


Classification + Localization

● Classification:
○ Input: Image
○ Output: Class label
○ Loss: Cross entropy (Softmaxlog)
○ Evaluation metric: Accuracy
● Localization:
○ Input: Image
○ Output: Box in the image (x, y, w, h)
○ Loss: L2 Loss (Euclidean distance)
○ Evaluation metric: Intersection over Union
● Classification + Localization:
○ Input: Image
○ Output: Class label + box in the image
○ Loss: Sum of both losses
Classification + Localization: ImageNet Challenge
● Dataset
○ 1000 Classes.
○ Each image has 1 class with at least one
bounding box.
○ ~800 Training images per class.
● Evaluation
○ Algorithm produces 5 (class + bounding box)
guesses.
○ Example is correct if at least one of guess has
correct class AND bounding box at least 50%
intersection over union.
Intersection Over Union (IoU)

● Important measurement for object localization.


● Used in both training and evaluation.

Intersection(A,B)
IoU(A,B)
= Union(A,B)
Classification + Localization: Model

Classification Head:
● C Scores for C
classes

Localization Head:
● Class agnostic:
(x,y,w,h)
● Class specific:
(x,y,w,h) X C
Computer Vision Tasks

Source: CS231n Object detection http://cs231n.stanford.edu/slides/2016/winter1516_lecture8.pdf


Object Detection 2001-2007
● Rapid Object Detection using a Boosted Cascade of
Simple Features (2001)
○ Viola & Jones
● Histograms of Oriented Gradients for Human
Detection (2005)
○ Dalal & Triggs
● Object Detection with Discriminatively Trained Part
Based Models (2010)
○ Felzenszwalb, Girshick, Ramanan
● Fast Feature Pyramids for Object Detection (2014)
○ Dollar
Object Detection 2007-2012

Source: Ross Girshick’s CVPR 2017 Tutorial http://deeplearning.csail.mit.edu/instance_ross.pptx


Object Detection Today

Source: Ross Girshick’s CVPR 2017 Tutorial http://deeplearning.csail.mit.edu/instance_ross.pptx


Object Detection: Datasets
2007 2013 2015
Pascal VOC ImageNet ILSVRC MS COCO
● 20 Classes ● 200 Classes ● 80 Classes
● 11K Training images ● 476K Training images ● 200K Training images
● 27K Training objects ● 534K Training objects ● 1.5M Training objects

Was de-facto standard, Essentially scaled up version More categories and more
currently used as quick of PASCAL VOC, similar object object instances in every
benchmark to evaluate new statistics. image. Only 10% of images
detection algorithms. contain a single object
category, 60% in Pascal. More
small objects than large
objects.
Pascal Examples
COCO Examples
Object Detection
● Input: Image
● Output: For each object class c and each
image i, an algorithm returns predicted
detections: locations with
confidence scores .
Object Detection: Evaluation

● True positive: correct class prediction AND IoU > 50%.


● False positive: wrong class or IoU < 50%.
● False negative: missed (not detected) object
● Only one detection can be matched to an object.

●
Object Detection: Evaluation

● Mean Average Precision (mAP) across all classes, based on Average Precision
(AP) per class, based on Precision and Recall.
Precision And Recall For a Threshold
Precision-Recall Curve

Source: Drawing by Prof. William H. Press, the University of Texas at Austin


Average Precision (AP)

● [In the vision community] AP is the estimated area under the PR curve

 
Mean Average Precision (mAP)

● The winner of each object class is the team with the highest average precision
● The winner of the challenge is the team with the highest mean Average
Precision (mAP) across all classes.
Object Detection: Evaluation

● Mean Average Precision (mAP) across all classes, based on Average Precision
(AP) per class, based on Precision and Recall.
Object Detection: Evaluation

● Today new metrics are emerging


○ Averaging precision over all IoU thresholds: 0.5:0.05:0.95
○ Averaging precision for different object sizes (small, medium, big)
○ Averaging recall as a metric to measure object proposal quality.
Looking for brilliant researchers

cv@brodmann17.com

You might also like