SSD Data Generator in Keras about advanced-deep-learning-with-keras HOT 4 CLOSED

tracing back my code, I believe I may have to correct myself. self.n_boxes is the total number of anchor boxes for all feature layers. The code for get_n_boxes() may be misleading which got me confused since each feature_shape is in fact feature_height x feature_width x n_anchors x 4 . 4 came from the number of offsets (there are 4 corners of a rectangle). The total number of anchor boxes per feature layer is (feature_height x feature_width x n_anchors x 4)/4. Since the number of anchors per feature pt is also 4, the computation was correct by accident.
in the data_generation, the inner for loop, generates the ground truth class, offset and mask per feature layer anchors. so, when you get the total, it is equal to self.n_boxes.

from advanced-deep-learning-with-keras.

Pls note that self.n_boxes refers to the number of bounding boxes per feature layer. In other words, if a feature layer is divided into 3x2 region, there are 6 bounding boxes. An anchor box is assigned to a bounding box as ground truth if: 1) it has max iou with that bounding box or 2) it has iou with that bounding box that is greater than a threshold (eg 0.6). Majority of these self.n_boxes bounding boxes are never assigned an anchor box (ie it is a background).

from advanced-deep-learning-with-keras.

Thank you for the explanation and clarification. However i still dont get two things in data_generator.py:

1. In the method get_n_boxes from class DataGenerator, as you mention, self.n_boxes is the sum of the number of bounding boxes (without taking into account the anchor boxes) for each feature layer:
   self.n_boxes += np.prod(shape) // self.n_anchors

Let's say that the total number of bounding boxes accross all the feature layers is 500 for example.

2 )Later on, in function data_generation, we create a tensor for storing class ground truth (line 130) which has dimensions gt_class.shape = (self.args.batch_size, self.n_boxes, self.n_classes), i.e, gt_class.shape = ( batch_size, 500, n_classes), if we consider 500 bounding boxes and 4 layers (1 for the background).

3. In line 160 in data_generation, we generate the anchors for each feature layer and feed it to the function get_gt_data to retrieve class ground truth tensor gt_cls for each layer. We append all the gt_cls across all the fature layers and get a tensor cls that has a shape of (self.n_boxes *4, n_classes).

4. Finally, we store this cls in the corresponding image at gt_class in line 182:
   gt_class[i] = cls

This is the question: How can we assign cls to gt_class if they have different shape? Whereas gt_class has 500 rows, cls has 2000 rows (since it takes into account the anchors per feature map point.

I have not run the code yet since I want to understand it before running it, and this difference in shape between gt_class and gt_cls is preventing me from advancing since I'm afraid I'm missing something important.

Thank you very much for your patience and help. I'm loving your book!
Pedro

from advanced-deep-learning-with-keras.

Thank you very much for your answer and for solving the question. I'll close the issue now.

from advanced-deep-learning-with-keras.