jaidevd / chart-ann Goto Github PK

Model sharing

• share the model so others can benefit. What's the best way to share it (Kaggle etc.)?

• add overflow-auto class to div.card-body.
• suggest previously added names (that aren't in the datalist) in the options.
• link instructions spreadsheet

The chart detection pipeline:

• windowing the input image
• classifying each window
• detecting neighbouring windows of the same class
• drawing bounding boxes around them

can theoretically happen in a single keras.models.Model object.

Currently, there are too many cpu -> gpu -> cpu ops happening in the pipeline since there are parts that are done with skimage and the rest with tf.keras.

But tensorflow supports

• Image windowing
• Image dilation
• Image erosion

So theoretically the skimage morphology operations can be replicated in tensorflow.

Suppose we have a single chart contained in an image sized 448 by 448 px. Our model would then split this into 4 boxes, and return a score of zero for them all, since a cropped quarter of a chart is not likely to be enough to identify it.

This is what we should do: window by size 224 * 224, do the usual run, but also by all higher multiples of 224, and aggregate the scores. Build a different heat map for each window size.

This will make it significantly slower, but hopefully more accurate.

in /validate, a barchart is mapped as dendrogram.

needs to be changed in the DB.