Part 2 Lesson 9 wiki

Part 2 & Alumni (2018)
134

What does SSD loss stand for again?

135

Single Shot Detector

3 Likes
136

so whatever classes are present it will only detect those , how about if a background image consists of an object which is not present in the classes , may be in that case softmax would take a leap forward?

137

For the multi label classification, why aren’t we multiplying the categorical loss by a constant like we did before? Before we multiplied categorical loss by 20 when we added it to bounding box loss.

5 Likes
138

why do we need even more bounding boxes, when one problem is some items (e.g. potted plant) are much larger than the anchor boxes already?

1 Like
139

we need bounding boxes of different sizes and different aspect ratios

3 Likes
140

okay, got it - thanks

141

anybody else getting lost in all of this or is it only me?

11 Likes
142

So is the Jaccard index between the Bounding box and the Anchor box?

1 Like
143

Yeah, I’m lost too. I don’t understand why we did all that effort to figure out which anchor box the bounding boxes belonged to. I thought we were saying that the anchor box was going to be the bounding box, but then bounding box came in from somewhere (from training, I guess), but… then… why were we doing anchor boxes?

5 Likes
144

Bounding boxes from ‘ground truth’ are bounding boxes from the training data.
Anchor boxes are our way of “guessing” where everything is assuming we don’t know where the objects are, as in the case of testing (and production).

Right?

1 Like
146

I think Jeremy didn’t really explain why the anchorboxes are needed to detect multiple objects

147

aren’t we modifying the anchor boxes (size, placement) to be bounding boxes?

3 Likes
148

2 reasons:
[1] you can avoid prediction clashes with more anchor boxes
[2] aspect ratios can help with IOU & clashes (person is vertical / motorcycle is horizontal - so the vertical and horizontal aspect ratios make it easy for the model)

2 Likes
149

Guess, It’ll take couple rewatch to grasp all that xD

2 Likes
150

I’m definitely going to need to put some time in this week to figure it all out.

151

There are a lot of pieces to keep track of, so this is material you’ll need to go back over a few times. It’s more helpful to have specific questions to ask during the lesson.

11 Likes
152

I think they only map the dataset so we can assign the SSD loss properly. It doesn’t look like a classifier for the (anchor) boxes.

153

Rewriting it is probably the only way.

154

And… why do we need to know what anchor box the object is in?

1 Like