1 / 51

Week 8 : Web-Assisted Object Detection

Week 8 : Web-Assisted Object Detection. Alejandro Torroella & Amir R. zamir. Geometry Method procedure. For each query image we manually set orientation, angle of view, range of view, and location of camera. : Camera location. : Object locations. : Field of vision.

falala
Télécharger la présentation

Week 8 : Web-Assisted Object Detection

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Week 8:Web-Assisted Object Detection Alejandro Torroella & Amir R. zamir

  2. Geometry Method procedure • For each query image we manually set orientation, angle of view, range of view, and location of camera. : Camera location : Object locations : Field of vision

  3. Geometry Method procedure Using the obtained FOV, select only the objects that are within the FOV Calculate the degrees from the left limit of the FOV and store in a vector specific to the object’s class These vectors will be our “true” layout of objects. . . .

  4. Geometry Method procedure We then run our DPM detectors for the classes in question on the query image. Below are results for Street Lights (green) and Traffic Signals (red).

  5. Geometry Method procedure We sift through the detections that completely disagree with the “true” GIS layout.

  6. Geometry Method procedure We then sift through the detections again by size of bounding boxes (too large or too small)

  7. Geometry Method procedure Using the sifted bounding boxes we generate all possible combinations (no repeats, order doesn’t matter) of possible layouts. For each class: Out of the N detections, choose k of them for the possible layout. Where For each layout combination we calculate the “cost” of it compared to the obtained “true” GIS layout and keep track of the combination that returned the minimum Two cost functions we’ve tested: Absolute value: Standard deviation:

  8. Geometry Method procedure Once we’ve traversed through all the possible combinations, we display the detections that resulted in the minimum of the cost function.

  9. Geometry Method results: Before Traffic Signals Trash Cans Traffic Signs

  10. Geometry Method results: After GIS Sift Traffic Signals Trash Cans Traffic Signs

  11. Geometry Method results: After Size Sift Traffic Signals Trash Cans Traffic Signs

  12. Geometry Method results: After Fusion (abs) Traffic Signals Trash Cans Traffic Signs

  13. Geometry Method results: after Fusion (std) Traffic Signals Trash Cans Traffic Signs

  14. Geometry Method results: Before Traffic Signals Street Lights

  15. Geometry Method results: After GIS Sift Traffic Signals Street Lights

  16. Geometry Method results: After Size Sift Traffic Signals Street Lights

  17. Geometry Method results: After Fusion (abs) Traffic Signals Street Lights

  18. Geometry Method results: after Fusion (std) Traffic Signals Street Lights

  19. Geometry Method results: Before Street Lights Traffic Signs

  20. Geometry Method results: After GIS Sift Street Lights Traffic Signs

  21. Geometry Method results: After Size Sift Street Lights Traffic Signs

  22. Geometry Method results: After Fusion (abs) Street Lights Traffic Signs

  23. Geometry Method results: after Fusion (std) Street Lights Traffic Signs

  24. Geometry Method results: Before Traffic Signals Street Lights

  25. Geometry Method results: After GIS Sift Traffic Signals Street Lights

  26. Geometry Method results: After Size Sift Traffic Signals Street Lights

  27. Geometry Method results: After Fusion (abs) Traffic Signals Street Lights

  28. Geometry Method results: after Fusion (std) Traffic Signals Street Lights

  29. Geometry Method results: Before Traffic Signals Street Lights

  30. Geometry Method results: After GIS Sift Traffic Signals Street Lights

  31. Geometry Method results: After Size Sift Traffic Signals Street Lights

  32. Geometry Method results: After Fusion (abs) Traffic Signals Street Lights

  33. Geometry Method results: after Fusion (std) Traffic Signals Street Lights

  34. Geometry Method results: Before Fire Hydrants Street Lights

  35. Geometry Method results: After GIS Sift Fire Hydrants Street Lights

  36. Geometry Method results: After Size Sift Fire Hydrants Street Lights

  37. Geometry Method results: After Fusion (abs) Fire Hydrants Street Lights

  38. Geometry Method results: after Fusion (std) Fire Hydrants Street Lights

  39. Geometry Method results: Before Street Lights Traffic Signs

  40. Geometry Method results: After GIS Sift Street Lights Traffic Signs

  41. Geometry Method results: After Size Sift Street Lights Traffic Signs

  42. Geometry Method results: After Fusion (abs) Street Lights Traffic Signs

  43. Geometry Method results: after Fusion (std) Street Lights Traffic Signs

  44. Geometry Method results: Before Traffic Signals Traffic Signs

  45. Geometry Method results: After GIS Sift Traffic Signals Traffic Signs

  46. Geometry Method results: After Size Sift Traffic Signals Traffic Signs

  47. Geometry Method results: After Fusion (abs) Traffic Signals Traffic Signs

  48. Geometry Method results: after Fusion (std) Traffic Signals Traffic Signs

  49. Geometry method Conclusions • Using the Standard deviation cost function resulted in better results compared to the Absolute value function on average. • A more advanced cost function would probably result in even better results • Results look promising considering we haven’t implemented a robust sensor model for creating the “true” GIS layout

  50. Goals for next week • Implement a robust sensor model • Look into more advanced cost functions • Instead of crudely sifting through the bounding boxes by size using a threshold based on the size of the image, use the distances of the objects from the camera to estimate how large the bounding box should be.

More Related