Different equation compared to paper? about squeezedet HOT 2 CLOSED

@shuuki4 thanks for your question. I guess your question is already answered in #4 . Equation (3) in the paper should be fixed, the code implementation is correct. Sorry for the confusion.

from squeezedet.

@BichenWuUCB I have one more question related to this issue. If the code implementation is correct, I think this part of the code (Line 170 of nn_skeleton.py) should be changed:

        box_center_x = tf.identity(
            anchor_x + delta_x * anchor_w, name='bbox_cx')
        box_center_y = tf.identity(
            anchor_y + delta_y * anchor_h, name='bbox_cy')
        box_width = tf.identity(
            anchor_w * util.safe_exp(delta_w, mc.EXP_THRESH),
            name='bbox_width')
        box_height = tf.identity(
            anchor_h * util.safe_exp(delta_h, mc.EXP_THRESH),
            name='bbox_height')

Since delta is divided by ground truth bboxes' width and height, gt boxes' width and height should be used when stretching box_center_x and box_center_y. However, this code uses anchor_w and anchor_h to rescale the box center. Therefore, if the code implementation is correct, this code should also be changed to use ground truth box width/height. Even if this information does not directly affect the learning process, it affects final prediction for raw bounding box coordinates and also affect visualizations, so I think this code should be fixed.

One more concern is that eval.py uses this code (model.det_boxes) for final bounding box prediction. Since there is no ground truth for test data split, it is impossible to rescale delta into real coordinates.

from squeezedet.