📗 -> Lecture Date: Name


Slides

🎤 Vocab

❗ Unit and Larger Context

Small summary

✒️ -> Scratch Notes

Object Detection

single Object?

-> Treat localization as a regression problem!

Multiple Objects?

  • Each image needs a different number of outputs
    -> Apply a CNN to many different crops of the image, CNN classifies each crop as object
    or background

Region Selection

In order to not have to do a CNN for a massive amount of potential crops (for multiple object search)…
-> find “blobby” regions likely to contain objects

R-CNN

Make CNN do roposas
Region Proposal Network (RPN) to redict proposals from features

  • Otherwise same as Fast R-CNN
    • Crop features for each proposal, classify each one
    • Basically train a classifier to detect object

Bipartite: 2 sets of N objects, cost between association object i to j has a cost.

  • Stable Marriage: Each object in each set tied to 1 and only one 1 object in the other set
    • Find the marriage that minimizes cost, make it stable
    • Solution to this is the hungarian algorithm, O(n)

R-CNN Benchmarks for Test-Time Speed:

  • R-CNN: 49
  • SPP-Net: 4.3
  • Fast R-CNN: 2.3
  • Faster R-CNN: 0.2

Single-Stage Object Detectors: YOLO / SSD / RetinaNet

Input image

  • 3 x H x W
    Divide image into grid
  • 7 x 7
  • Image a set of base boxes centered at each grid cell
    • Here B = 3
      Within each grid cell:
  • Regress from each of the B base boxes to a final box with 5 numbers:
    • (dx, dy, dh, dw, confidence)
  • Predict scores for each of C classes (including background as a class)
  • Looks a lot like RPN, but category-specific!
    Output:
  • 7 x 7 x (5 * B + C

Mask R-CNN (looks to identify something and give a mask outline (countour kind of))
Also does pose

Object Detection + Captioning = Dense Captioning

  • Can extrapolate out to do dense video captioning

Objects + Relationships = Scene Graphs

  • Captures relationships between objects
    • Painting is hanging on a wall
      • Painting -> Hanging -> Wall
    • Very useful for robotics

Resources

  • Put useful links here

Connections

  • Link all related words