Cost function in Yolo-v1

[christophesaintjean]

Hi, I think there is an error with the cost function in Yolo-v1. In the original paper, the authors said that the confidence value should be :

• the IOU for the best box in a cell that contains an object (1_{ij}^{obj})
• and zero elsewhere (1_{ij}^{noobj}) For me, the confidence of boxes in a cell that contains an object BUT are not the best should be 0. This leads to a different formulation of self['iou_normal'] which appears hard to reproduce without model.confidence variable: self['iou_normal'] = tf.reduce_sum(mask_normal * tf.square(model.confidence), name='iou_normal') Do you think i am right or wrong ? Best regards, Christophe.

[TaihuLight]

The cost function in YOLO-V2 is right in this project?

[ruiminshen]

@christophesaintjean Thank you for your question. In the function model.yolo.Objectives.init, the tensors "mask_best" and "masknormal" representing $1{ij}^{obj}$ and $1_{ij}^{noobj}$, respectively.

The tensor "mask_best" requires two conditions: the cell contains an object (self.mask) AND the bbox has the best IoU value in its cell (best_box). Because "best_box_iou" calculates the best IoU value of each independent cell, and "best_box" requires the IoU value of a bbox equals "best_box_iou". So it will be 0 if its IoU is not the best in its cell.

[ruiminshen]

@TaihuLight Yes, I've checked it.

[christophesaintjean]

@ruiminshen, i studied more your code and noticed that i agree your comment I explain it below for the interested reader.

• best_box_iou : best IOU whatever the cell (even if no object is the corresponding cell)
• bestbox : (tensor) indicator for being the best box_
• maskbest : (tensor) indicator for being "the best box and in a cell that contains an object" $= 1{ij}^{obj}$
• masknormal (tensor) indicator for being "not the best box or in a cell that doesn't contain an object" $ = 1{ij}^{noobj}$
• iou_dist = tf.square(model.iou - mask_best, name='iou_dist')
  • for boxes when $1_{ij}^{obj} = 1, iou_dist = ||model.iou - 1||^2$
  • for boxes when $1_{ij}^{obj} = 0, ioudist = ||model.iou - 0||^2$ , they are not responsible for prediction_
• cnt = np.multiply.reduce(iou_dist.get_shape().as_list()) new term in this version ? numbers of boxes in whole batch, true ?
• self['iou_best'] = tf.identity(tf.reduce_sum(mask_best * iou_dist) / cnt, name='iou_best')
  • Now this loss includes an expectation over all boxes
  • for boxes when 1{ij}^{obj} = 1, loss is E[||iou - 1||^2] -> push them to 1 (the ideal IOU)_
  • for boxes when 1{ij}^{obj} = 0, loss is 0 # so no influence_
• self['iou_normal'] = tf.identity(tf.reduce_sum(mask_normal * iou_dist) / cnt, name='iou_normal')
  • for boxes when $1{ij}^{obj} = 1$, loss is 0, so no influence
  • for boxes when $1_{ij}^{obj} = 0$, loss is $E[||iou||^2]$ -> push them to 0 (no confidence)
  • This is what i called ' tf.square(model.confidence)'

I am implementing Yolo-v1 with Keras. My implementation for these two losses are as the following:

• confidence_positive_loss = K.sum(one_ijobj * K.square(iou - C)) # or K.mean(...)
• confidence_negative_loss = K.sum(one_ijnoobj * K.square(C)) where IOU is computed as in your code and C_ is the confidence output but the model. This latter is a tensor [batch_size, nb_cells, nb_boxes] in your code.

So maybe, there is a very subtle difference between our interpretations:

• I want the model confidence for the best box to be equal to the IOU (and IOU is maximized through coordinates loss)
• you want the model confidence for the best box to be equal to 1

At the end, it is the same loss since IOU is 1 for the best box in the learning step. I got this because i have encoded the image annotation into a nb_cells(nb_classes + boxes_per_cell (1 + 4)) vector. However, only the first box in each cell is used for the desired output.

Best regards, Christophe

ps: thank you very much for having shared your valuable code.