![Page 1: Carl Vondrick, Aditya Khosla, Tomasz Malisiewicz, Antonio Torralba Massachusetts Institute of Technology](https://reader033.vdocument.in/reader033/viewer/2022051104/5a4d1b6e7f8b9ab0599b4565/html5/thumbnails/1.jpg)
HOGgles: Visualizing Object Detection Features
Carl Vondrick, Aditya Khosla, Tomasz Malisiewicz, Antonio Torralba
Massachusetts Institute of Technologyfvondrick,khosla,tomasz,[email protected]
![Page 2: Carl Vondrick, Aditya Khosla, Tomasz Malisiewicz, Antonio Torralba Massachusetts Institute of Technology](https://reader033.vdocument.in/reader033/viewer/2022051104/5a4d1b6e7f8b9ab0599b4565/html5/thumbnails/2.jpg)
Introduction Feature Visualization Algorithms Evaluation of Visualizations Conclusion
OUTLINE
![Page 3: Carl Vondrick, Aditya Khosla, Tomasz Malisiewicz, Antonio Torralba Massachusetts Institute of Technology](https://reader033.vdocument.in/reader033/viewer/2022051104/5a4d1b6e7f8b9ab0599b4565/html5/thumbnails/3.jpg)
Figure 1 shows a high scoring detection from an object detector with HOG features and a linear SVM classifier trained on PASCAL.
Introduction
![Page 4: Carl Vondrick, Aditya Khosla, Tomasz Malisiewicz, Antonio Torralba Massachusetts Institute of Technology](https://reader033.vdocument.in/reader033/viewer/2022051104/5a4d1b6e7f8b9ab0599b4565/html5/thumbnails/4.jpg)
Figure 2 shows the output from our visualization on the features for the false car detection.
Introduction
![Page 5: Carl Vondrick, Aditya Khosla, Tomasz Malisiewicz, Antonio Torralba Massachusetts Institute of Technology](https://reader033.vdocument.in/reader033/viewer/2022051104/5a4d1b6e7f8b9ab0599b4565/html5/thumbnails/5.jpg)
Figure 3 inverts more top detections on PASCAL for
a few categories
Introduction
![Page 6: Carl Vondrick, Aditya Khosla, Tomasz Malisiewicz, Antonio Torralba Massachusetts Institute of Technology](https://reader033.vdocument.in/reader033/viewer/2022051104/5a4d1b6e7f8b9ab0599b4565/html5/thumbnails/6.jpg)
Figure 4 for a comparison between our visualization and HOG glyphs.
Introduction
![Page 7: Carl Vondrick, Aditya Khosla, Tomasz Malisiewicz, Antonio Torralba Massachusetts Institute of Technology](https://reader033.vdocument.in/reader033/viewer/2022051104/5a4d1b6e7f8b9ab0599b4565/html5/thumbnails/7.jpg)
We pose the feature visualization problem as one of feature inversion.
be an image and be the corresponding HOG feature descriptor.
Feature Visualization Algorithms
![Page 8: Carl Vondrick, Aditya Khosla, Tomasz Malisiewicz, Antonio Torralba Massachusetts Institute of Technology](https://reader033.vdocument.in/reader033/viewer/2022051104/5a4d1b6e7f8b9ab0599b4565/html5/thumbnails/8.jpg)
Baseline A: Exemplar LDA (ELDA)
Consider the top detections for the exemplar object detector for a few images shown in Figure 5
Feature Visualization Algorithms
![Page 9: Carl Vondrick, Aditya Khosla, Tomasz Malisiewicz, Antonio Torralba Massachusetts Institute of Technology](https://reader033.vdocument.in/reader033/viewer/2022051104/5a4d1b6e7f8b9ab0599b4565/html5/thumbnails/9.jpg)
We use this simple observation to produce our first inversion baseline
Suppose we wish to invert HOG feature y.
We first train an exemplar LDA detector for this query
Feature Visualization Algorithms
![Page 10: Carl Vondrick, Aditya Khosla, Tomasz Malisiewicz, Antonio Torralba Massachusetts Institute of Technology](https://reader033.vdocument.in/reader033/viewer/2022051104/5a4d1b6e7f8b9ab0599b4565/html5/thumbnails/10.jpg)
The HOG inverse is then the average of the top K detections in RGB space
Feature Visualization Algorithms
![Page 11: Carl Vondrick, Aditya Khosla, Tomasz Malisiewicz, Antonio Torralba Massachusetts Institute of Technology](https://reader033.vdocument.in/reader033/viewer/2022051104/5a4d1b6e7f8b9ab0599b4565/html5/thumbnails/11.jpg)
Baseline B: Ridge Regression We present a fast, parametric inversion baseline
based off ridge regression.
Feature Visualization Algorithms
![Page 12: Carl Vondrick, Aditya Khosla, Tomasz Malisiewicz, Antonio Torralba Massachusetts Institute of Technology](https://reader033.vdocument.in/reader033/viewer/2022051104/5a4d1b6e7f8b9ab0599b4565/html5/thumbnails/12.jpg)
In order to invert a HOG feature y, we calculate the most likely image from the conditional Gaussian distribution
Feature Visualization Algorithms
![Page 13: Carl Vondrick, Aditya Khosla, Tomasz Malisiewicz, Antonio Torralba Massachusetts Institute of Technology](https://reader033.vdocument.in/reader033/viewer/2022051104/5a4d1b6e7f8b9ab0599b4565/html5/thumbnails/13.jpg)
Baseline C: Direct Optimization We now provide a baseline that attempts to find
images that, when we compute HOG on it, sufficiently match the original descriptor.
Let
Feature Visualization Algorithms
![Page 14: Carl Vondrick, Aditya Khosla, Tomasz Malisiewicz, Antonio Torralba Massachusetts Institute of Technology](https://reader033.vdocument.in/reader033/viewer/2022051104/5a4d1b6e7f8b9ab0599b4565/html5/thumbnails/14.jpg)
3.4. Algorithm D: Paired Dictionary Learning
Feature Visualization Algorithms
![Page 15: Carl Vondrick, Aditya Khosla, Tomasz Malisiewicz, Antonio Torralba Massachusetts Institute of Technology](https://reader033.vdocument.in/reader033/viewer/2022051104/5a4d1b6e7f8b9ab0599b4565/html5/thumbnails/15.jpg)
Inversion Benchmark
Evaluation of Visualizations
![Page 16: Carl Vondrick, Aditya Khosla, Tomasz Malisiewicz, Antonio Torralba Massachusetts Institute of Technology](https://reader033.vdocument.in/reader033/viewer/2022051104/5a4d1b6e7f8b9ab0599b4565/html5/thumbnails/16.jpg)
Visualization Benchmark
Evaluation of Visualizations
![Page 17: Carl Vondrick, Aditya Khosla, Tomasz Malisiewicz, Antonio Torralba Massachusetts Institute of Technology](https://reader033.vdocument.in/reader033/viewer/2022051104/5a4d1b6e7f8b9ab0599b4565/html5/thumbnails/17.jpg)
Evaluation of Visualizations
![Page 18: Carl Vondrick, Aditya Khosla, Tomasz Malisiewicz, Antonio Torralba Massachusetts Institute of Technology](https://reader033.vdocument.in/reader033/viewer/2022051104/5a4d1b6e7f8b9ab0599b4565/html5/thumbnails/18.jpg)
Evaluation of Visualizations
![Page 19: Carl Vondrick, Aditya Khosla, Tomasz Malisiewicz, Antonio Torralba Massachusetts Institute of Technology](https://reader033.vdocument.in/reader033/viewer/2022051104/5a4d1b6e7f8b9ab0599b4565/html5/thumbnails/19.jpg)
We believe visualizations can be a powerful tool for
understanding object detection systems and advancing research in computer vision.
Since object detection researchers analyze HOGglyphs everyday and nearly every recent object detection paper includes HOG visualizations
Conclusion