Downloaded 482 times
Uploaded byEntrepreneur / Startup
You only look once (YOLO) : unified real time object detection
YOLO (You Only Look Once) is a real-time object detection system that frames object detection as a regression problem. It uses a single neural network that predicts bounding boxes and class probabilities directly from full images in one evaluation. This approach allows YOLO to process images and perform object detection over 45 frames per second while maintaining high accuracy compared to previous systems. YOLO was trained on natural images from PASCAL VOC and can generalize to new domains like artwork without significant degradation in performance, unlike other methods that struggle with domain shift.
In this document
Powered by AI
Overview of YOLO (You Only Look Once) approach to real-time object detection by key researchers.
Comparison of previous methods like DPM and R-CNN with YOLO's approach involving convolutional networks.
YOLO treats object detection as a regression task using a single neural network for speed and accuracy.
YOLO performs feature extraction and bounding box predictions through a grid-based method.
Each grid cell predicts multiple bounding boxes and confidence scores based on IOU with ground truth.
Prediction of class probabilities alongside bounding box coordinates, utilizing thresholding techniques.
Model predictions are encoded in a tensor format incorporating bounding box and class probabilities.
Detailed breakdown of the neural network design inspired by GoogLeNet for effective image classification.
Steps involved in training YOLO on the ImageNet dataset and PASCAL VOC with parameters for optimal results.
Describes the inference stage in YOLO, with emphasis on bounding box predictions and non-maximal suppression.
Identifies limitations of YOLO, particularly in detecting small objects and handling various object ratios.
Comparative analysis of YOLO against other real-time detection systems highlighting efficiency and design.
Results of YOLO experiments on PASCAL VOC 2007, including real-time performance metrics.
Exploration of combining YOLO with Fast R-CNN, noting improved detection with potential speed trade-off.
Assessment of YOLO's performance in detecting small objects and the combined efficiency with R-CNN.
YOLO's effectiveness in detecting people in artwork, showcasing its robust generalizability.
Summary of YOLO results on various images demonstrating its object detection capabilities.
More Related Content
![[PR12] You Only Look Once (YOLO): Unified Real-Time Object Detection](https://cdn.slidesharecdn.com/ss_thumbnails/yolo-170616085751-thumbnail.jpg?width=640&height=640&fit=bounds)
Similar to You only look once (YOLO) : unified real time object detection
You only look once (YOLO) : unified real time object detection
- 1. You Only LookOnce (YOLO): Unified Real-Time Object Detection Joseph Redmon, Santosh Divvala, Ross Girshick, Ali Farhadi University of Washington, Allen Institute for AI, Facebook AI Research ~ Ashish
- 2. Previously : ObjectDetection by Classifiers ● DPM (Deformable Parts Model) ○ Sliding window → classifier (evenly spaced locations) ● R-CNN ○ Region proposal --> potential BB ○ Run classifiers on BB ○ Post processing (refinement, eliminate, rescore) ● YOLO ○ Resize image, run convolutional network, non-max suppression
- 3. YOLO : ObjectDetection as Regression Problem ● output: Bounding box coordinates and Class Probabilities ● Single Neural Network ● Benefits: ○ Extremely Fast (one NN + 45 frames per sec), twice more mAP. ○ Global Reasoning (knows context, less background errors) ○ Generalizable Representations (train natural images, test art-work, applicable new domain)
- 4. Unified Detection ● FeatureExtraction ○ Predict all class BB simultaneously ● SxS Grid ○ Each cell predicts B bounding boxes + Confidence Score ● Confidence Score ○ Confidence is IOU between predicted box and any ground truth box = ● Class Probability ● Tensor
- 5. Detection Process (YOLO)Grid SXS S = 7
- 6. Confidence Score Each gridcell predicts B bounding boxes and confidence scores for those boxes. If a cell has an object , then confidence score = Intersection over union (IOU) between the predicted box and the ground truth.
- 7. Detection Process (YOLO) Eachcell predicts B boxes(x,y,w,h) and confidences of each box: P(Object) .(x,y) w h B = 2 Prob. that box contains an object P1, P2 No Object
- 8. Each cell predictsBounding Boxes and Confidence .(x,y)
- 9. Each cell alsopredicts class probability Bicycle Dog Car E.g. Dog : 0.8 Car : 0 Bicycle : 0 E.g. Dog : 0 Car : 0 Bicycle : 0.7 E.g. Dog : 0 Car : 0.7 Bicycle : 0
- 10. Bounding Boxes +Class Prediction .(x,y) P (class) = P (class|object) x P(object) Thresholding
- 11. Model These predictions areencoded as Tensor of dimension (SxSx(Bx5+C)) SxS grid, C = class probability, B= no of bounding boxes.
- 12. Network Design ● Inspiredby the GoogLeNet (image classification) ● 24 convolutional layers followed by 2 fully connected layers ● Fast YOLO uses 9 convolutional layers (instead of 24)
- 13. Training 1. Pretrain onImageNet 1000 dataset 2. 20 convolutional layers + an average pooling layer + a fully connected layer 3. Trained for 1 week, accuracy 88% (ImageNet 2012 validation dataset) 4. Convert model to perform detection 5. Added 4 convolutional layer + 2 fully connected layer + increased input resolution from 224 x 224 to 448 x 448. 6. Final layer predicts class probabilities + BB. 7. Linear activation function (final layer), Relu (all other layers) 8. Sum of squared error as loss function (easy to optimise)
- 14.
- 15. Training - Validation 1.Train network for 135 epochs on the training and validation data sets from PASCAL VOC 2007 AND 2012 2. Testing data VOC 2007 & 2012 3. Batch size = 64, momentum = 0.9, decay = 0.0005 4. Learning rate : a. First few epochs , raise LR 10^-3 to 10^-2 b. Model diverges if starting LR is high due to unstable gradient c. first 75 epoch, LR 10^-2 d. next 30 epochs, LR 10^-3 e. next 30 epochs, LR 10^-4 5. To avoid overfitting: a. Dropout layer with rate 0.5 b. For Data Augmentation, scaling and translation up to 20% of original image size
- 16. Inference ● On PASCALVOC YOLO predicts 98 BB per image and class probability for each box. ● Objects near border are localised by multiple cells ○ Non Maximal suppression can be used to fix these multiple detections (Non-max suppression is a way to eliminate points that do not lie in important edges. ) ■ Adds 2 to 3% to mAP
- 17. Limitation of YOLO ●Struggle with small objects ● Struggles with difference aspects and ratio of objects ● Loss function treats error in different size of boxes same
- 18. Comparison with otherReal time Systems: ● DPM : disjoint pipeline (sliding window, features, classify, predict BB) - YOLO concurrently ● R-CNN : region proposal , complex pipeline ( predict bb, extract features, non-max suppression) - 40 sec per image (2000 BB) : YOLO : 98 BB ● Deep Multibox : cnn, cannot do general detection ● OverFeat : cnn, disjoint system, no global context ● MultiGrasp : similar in design (YOLO) , only find a region
- 19. Experiments ● PASCAL VOC 2007 ●Realtime : ○ YOLO VS DPM 30 Hz
- 20. VOC 2007 ErrorAnalysis
- 21. Combining Fast R-CNNand YOLO ● YOLO makes fewer background mistakes than Fast R-CNN ● This combination doesn’t benefit from the speed of YOLO since each model is run separately and then combine the results.
- 22. VOC 2012 Results ●YOLO struggles with small objects (bottle, sheep, tv/monitor) ● Fast R-CNN + YOLO : Highest performing detection methods
- 23. Generalizability: Person Detectionin Artwork ● YOLO has good performance on VOC 2007 ● Its AP degrades less than other methods when applied to artwork. ● Artwork / Natural Images are very different on a pixel level but very similar in terms of size and shape, so YOLO predicts good bounding boxes and detections.
- 24.
- 25. Darknet (YOLO) Resultson random images