Add predicted objects to predict command

Depends #35 (closed)

When we are able to get the coverage vector for each detected object, we can work on finding polygons for each of them. This will be integrated to the predict command and triggered by a new flag --predict-objects.

Implement two functions in dan.predict.attention

def get_predicted_polygons(
    text,
    weights,
    level,
    word_separators=["\n", " "],
    line_separators=["\n"],
):
    polygons = []
    # Split text into characters, words or lines
    text_list, offset = split_text(text, level, word_separators, line_separators)
    max_value = weights.sum(0).max()
    return [
        get_polygon(text_piece, level, max_value, weights)
        for text_piece in text_list
    ]

def get_polygon(text_piece, level, max_value, weights):
    coverage_vector = compute_coverage(text_piece, max_value, level, weights)

    # Generate a binary image for the current channel.
    bin_img = coverage_vector.copy()
    bin_img[bin_img > 0] = 1

    # Detect the objects contours.
    contours, _ = cv2.findContours(
        np.uint8(bin_img), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE
    )

    return {
        "confidence": ..., # average of coverage vector on contour
        "polygon": contour,
        "type": level,
    }

The get_polygon function will be used in dan.predict.attention.plot_attention if a new boolean parameter output_polygons is enabled. The loop in plot_attention becomes

polygons = []
for text_piece in text_list:
    coverage_vector = compute_coverage(...)
    if output_polygons:
         polygons.append(get_polygon(...))
    
    # Keep track of text length
    tot_len += len(text_piece) + offset

    attention_map.append(blend_coverage(...))
return polygons

I supposed that the image generation can also be ported in a separate blend_coverage function.

The get_predicted_polygons function will be used in the predict method when attentions is enabled. We will add a new parameter attention_level which will be used as level in the get_predicted_polygons call. The result will be stored in a new key objects.

The predict command will have a new CLI argument --predict-objects, this flag will control the output_polygons value in the plot_attention call