Download as PDF, PPTX
Uploaded byBrodmann17
Advanced deep learning based object detection methods
The document discusses advanced deep learning methods for object detection, focusing on improving non-maximum suppression (NMS), multi-scale detection, and the implementation of focal loss for better class imbalance handling. It highlights the transition from traditional NMS to learned approaches and introduces architectures such as Feature Pyramid Networks (FPN) and Mask R-CNN, emphasizing their efficiency and effectiveness in various object detection tasks. Additionally, it references practical benchmarks and external educational resources for deeper exploration of the topic.
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)
![Deformable DETR Review [CDM]](https://cdn.slidesharecdn.com/ss_thumbnails/deformabledetrreviewcdm-201113070345-thumbnail.jpg?width=640&height=640&fit=bounds)
Similar to Advanced deep learning based object detection methods
Advanced deep learning based object detection methods
- 1. Advanced Deep Learningbased Object Detection Methods
- 2. Improving Object DetectionWith One Line of Code ● Non-Maximum Suppression is a greedy process. ○ It worked well enough in 2007 but it doesn’t anymore. ● High scoring detections can be suppressed just as low scoring detections. ○ Overlap with stronger detection is the only criteria. ● Should one detection completely suppress another detection, or simply reduce its confidence?
- 3. Improving Object DetectionWith One Line of Code ● NMS: ● Linear Soft-NMS: ● Gaussian Soft-NMS: ○ Linear Soft-NMS is not continuous in terms of overlap and a sudden penalty is applied when a NMS threshold is reached. ○ Instead we can use a continuous function:
- 4. Improving Object DetectionWith One Line of Code
- 5. Improving Object DetectionWith One Line of Code
- 6. Learning Non-Maximum Suppression ●Object detectors are mostly trained end-to-end, except for the NMS. ○ NMS is still fully hand-crafted, and forces a trade-off between recall and precision. ● Training loss is not evaluation loss. ○ Training is performed without NMS ○ During evaluation, multiple detections for same object count as false positives. ● Instead, train the network to include the suppression process. ○ Only output one bounding box per object. ○ Learn how to handle close objects.
- 7. Learning Non-Maximum Suppression ●Additional blocks that: ○ Encode pairwise information. ○ For each detection, pool information from all pairings. ○ Update feature vector. ○ Repeat. ● New loss: ○ Only one positive candidate per object. ○ Instead of the current practice to take all objects with IoU>50%
- 8.
- 9.
- 10. ● Multi-scale objectdetection using image pyramid ○ Predict different scales by applying same model at different image resolutions. ● Classic method. ● But also, in OverFeat. ● Slow. Requires multiple evaluation of the same model. Multi-Scale Object Detection
- 11. Multi-Scale Object Detection ●Predict multiple scale of objects using a single feature map. ● Same as Faster R-CNN. ● Fast ● Single model (same in training as in testing). ● Bad features resolution for small objects.
- 12. ● Predict differentobject sizes at different feature scales. ● Same as SSD. ● Good features resolution for small objects ● But features are much weaker than in deeper layers. Multi-Scale Object Detection
- 13. ● Single model(same in training as in testing). ● Good features resolution for small objects. ● Strong features in all layers. ● Almost no overhead over SSD (= Fast). Feature Pyramid Network (FPN)
- 14.
- 15. Feature Pyramid Network(FPN) ● How important is top-down enrichment? ● How important are lateral connections? ● How important are pyramid representations?
- 16. Feature Pyramid Network(FPN) ● How important is top-down enrichment? ● How important are lateral connections? ● How important are pyramid representations?
- 17. Focal Loss forDense Object Detection ● Can we train a single stage detector to be as accurate as two stage detectors? ● Contributions: ○ RetinaNet: Single stage object detector based on FPN backbone. ○ New loss.
- 18. Focal Loss forDense Object Detection ● Class unbalance is an important issue for object detection. ● Previous solutions: ○ Random resampling at 1:3 ratio. ○ Hard negative resampling at 1:3 ratio. ● Both solutions means that at each step, we only a few samples actually matters to the loss function. ● Instead, include all samples but use different weight for each class. ○ Regular cross entropy: ○ Weighted cross entropy:
- 19. ● Using weightCE as baseline: ○ Can we do better? ○ Can we use different weight for each sample? ● Focal loss: ● Every sample is weighted according to its error. ○ We want to focus on samples which are mislabeled. Focal Loss for Dense Object Detection
- 20. ● Different parametersfor RetinaNet Focal Loss for Dense Object Detection
- 21. ● Comparison withonline hard negative mining Focal Loss for Dense Object Detection
- 22. ● Accuracy/speed trade-offs FocalLoss for Dense Object Detection
- 23. ● Benchmark results FocalLoss for Dense Object Detection
- 24. Also Read: Deformable ConvolutionalNetworks https://arxiv.org/abs/1703.06211
- 25. YouTube Videos ● CS231n ○Lecture 11 - Detection and segmentation https://youtu.be/nDPWywWRIRo ● Deep Learning for Objects and Scenes (CVPR 2017 Workshop) ○ Lecture 1: Learning Deep Representations for Visual Recognition, by Kaiming He https://youtu.be/jHv37mKAhV4 ○ Lecture 2: Deep Learning for Instance-level Object Understanding, by Ross Girshick https://youtu.be/jHv37mKAhV4?t=39m4s
- 26. Looking for brilliantresearchers [email protected] / [email protected]
- 27. Computer Vision Tasks Source:CS231n Object detection http://cs231n.stanford.edu/slides/2016/winter1516_lecture8.pdf
- 28. Mask R-CNN ● Instancesegmentation with pose estimation for people. ● Extends faster R-CNN by adding new branch for the instance mask task. ● Pose estimation can be added by simply adding an additional branch. ● SOTA accuracy on detection, segmentation and pose estimation at 5 FPS on GPU. ● https://arxiv.org/abs/1703.06870 ● Girshick won young researcher award.
- 29.
- 30.
- 31.
- 32. Mask R-CNN ● RoiPool ○Quantization breaks pixel-to-pixel alignment ○ Too coarse and not good for fine spatial information required for mask. ● RoiAlign ○ Bilinearly sample the proposal region and avoid the quantization. ○ Smoothly normalize features and predictions into coordinate frame free of scale and aspect ratio
- 33.
- 34. Mask R-CNN ● Backbonearchitecture ○ ResNet ○ ResNeXt ○ FPN ● Mask representation ○ FC vs. Convolutional ○ Multinomial vs. Independent Masks: softmax vs. sigmoid ○ Class-Specific vs. Class-Agnostic Masks: almost same accuracy ● Multi-task learning ○ Mask task improves object detection accuracy. ○ Keypoint task reduces object detection accuracy.
- 35. Mask R-CNN ● Poseestimation ○ Simply add an additional branch. ○ Model a keypoint’s location as a one-hot mask, and adopt Mask R-CNN to predict K masks. ○ Experiments are mainly to demonstrate the generality of the Mask R-CNN framework. ○ RoiAlign improves this task’s accuracy as well.
- 36. Looking for brilliantresearchers [email protected]