Skip to content

README.md

Latest commit

 

History

History
182 lines (104 loc) · 9.91 KB

README.md

File metadata and controls

182 lines (104 loc) · 9.91 KB
Raw

Object Detection

Di image classification models we don deal wit before na to take one image and produce one categorical result, like di class 'number' for MNIST problem. But, for many cases, we no just wan sabi say di picture get objects - we wan fit know di exact place wey dem dey. Na wetin object detection dey do be dat.

Pre-lecture quiz

Object Detection

Image from YOLO v2 web site

One Simple Way to Do Object Detection

If we wan find cat for one picture, one simple way to do object detection go be like dis:

  1. Cut di picture into plenty small tiles.
  2. Run image classification for each tile.
  3. Any tile wey get high activation fit mean say di object dey inside.

Naive Object Detection

Image from Exercise Notebook

But dis method no too make sense because e no dey fit locate di object bounding box well. To get di exact location, we go need to run regression to predict di bounding box coordinates - and for dat, we go need special datasets.

Regression for Object Detection

Dis blog post explain di basics of detecting shapes well well.

Datasets for Object Detection

You fit see dis datasets for dis kind task:

  • PASCAL VOC - 20 classes
  • COCO - Common Objects in Context. 80 classes, bounding boxes and segmentation masks

COCO

Object Detection Metrics

Intersection over Union

For image classification, e dey easy to measure how di algorithm dey perform. But for object detection, we need to measure di class correctness and di bounding box location precision. For di bounding box, we dey use Intersection over Union (IoU), wey dey measure how two boxes (or areas) dey overlap.

IoU

Figure 2 from dis blog post on IoU

Di idea simple - we go divide di area wey two figures dey overlap by di area of their union. If di two areas dey exactly di same, IoU go be 1. If dem no touch at all, IoU go be 0. Any other case go dey between 0 and 1. We dey usually consider bounding boxes wey get IoU above one certain value.

Average Precision

If we wan measure how well one class of objects $C$ dey recognized, we go use Average Precision metrics. Di steps be:

  1. Look di Precision-Recall curve wey show accuracy based on detection threshold (0 to 1).
  2. Di threshold go affect how many objects we detect and di precision/recall values.
  3. Di curve go look like dis:

Image from NeuroWorkshop

Di Average Precision for one class $C$ na di area under di curve. More specifically, Recall axis dey divided into 10 parts, and Precision dey averaged for all di points:

$$ AP = {1\over11}\sum_{i=0}^{10}\mbox{Precision}(\mbox{Recall}={i\over10}) $$

AP and IoU

We go only consider detections wey get IoU above one certain value. For example, for PASCAL VOC dataset, $\mbox{IoU Threshold} = 0.5$, but for COCO, AP dey measured for different $\mbox{IoU Threshold}$ values.

Image from NeuroWorkshop

Mean Average Precision - mAP

Di main metric for Object Detection na Mean Average Precision or mAP. E dey calculate di Average Precision across all object classes, and sometimes across $\mbox{IoU Threshold}$. You fit read more about how to calculate mAP for dis blog post) and dis one with code samples.

Different Object Detection Approaches

We get two main types of object detection algorithms:

  • Region Proposal Networks (R-CNN, Fast R-CNN, Faster R-CNN). Di idea na to generate Regions of Interests (ROI) and run CNN for dem to find maximum activation. E dey similar to di simple method, but di ROIs dey generated in a smarter way. Di problem be say e dey slow because we dey run CNN classifier many times for di image.
  • One-pass (YOLO, SSD, RetinaNet) methods. For dis ones, di network dey predict both classes and ROIs in one pass.

R-CNN: Region-Based CNN

R-CNN dey use Selective Search to generate ROI regions, wey CNN go process to extract features. SVM-classifiers go determine di object class, and linear regression go predict di bounding box coordinates. Official Paper

RCNN

Image from van de Sande et al. ICCV’11

RCNN-1

*Images from dis blog

F-RCNN - Fast R-CNN

Dis method dey similar to R-CNN, but e dey define regions after convolution layers don run.

FRCNN

Image from di Official Paper, arXiv, 2015

Faster R-CNN

Dis method dey use neural network to predict ROIs - di Region Proposal Network. Paper, 2016

FasterRCNN

Image from di official paper

R-FCN: Region-Based Fully Convolutional Network

Dis algorithm dey faster pass Faster R-CNN. Di main idea be:

  1. Extract features wit ResNet-101.
  2. Process di features wit Position-Sensitive Score Map. Each object from $C$ classes dey divide into $k\times k$ regions, and we dey train to predict di object parts.
  3. For each part, di network go vote for di object class, and di class wit di highest vote go dey selected.

r-fcn image

Image from official paper

YOLO - You Only Look Once

YOLO na realtime one-pass algorithm. Di main idea be:

  • Divide di image into $S\times S$ regions.
  • For each region, CNN go predict $n$ possible objects, bounding box coordinates, and confidence=probability * IoU.

YOLO

Image from official paper

Other Algorithms

✍️ Exercises: Object Detection

Continue your learning for dis notebook:

ObjectDetection.ipynb

Conclusion

For dis lesson, you don see di different ways wey object detection fit work!

🚀 Challenge

Read dis articles and notebooks about YOLO and try dem:

Post-lecture quiz

Review & Self Study

Assignment: Object Detection

Disclaimer:
Dis docu don use AI translation service Co-op Translator take translate am. Even though we dey try make e accurate, abeg sabi say automated translations fit get mistake or no dey 100% correct. Di original docu for di language wey dem write am first na di main correct one. For important information, e good make una use professional human translation. We no go fit take blame for any misunderstanding or wrong interpretation wey fit happen because of dis translation.