2015-12-14-notes-on-yolo.md

layout	category	title	date
post	deep_learning	Notes On YOLO	2015-12-14

How does YOLO organise training data?

ground-truth:

49 x (1 + 20 + 4) =>
49 x (1 x obj_gt + 20 x classes_gt + 4 x box_gt)

predict-data:

49 x 20 + 49x(1x2) + 49x(4x2) =>
49 x (20 x classes) + 49 x (2 x obj_confidence) + 49 x (2 x predict_boxes)

Some questions..

(1) The multi-part loss function differ from the code implementation:

$$ \lambda_{coord} \sum_{i=0}^{S^2} \sum_{j=0}^{S} \mathcal{1}_{ij}^{obj} (x_i - \hat{x}_i)^2 + (y_i - \hat{y}_i)^2 $$

$$

• \lambda_{coord} \sum_{i=0}^{S^2} \sum_{j=0}^{S} \mathcal{1}_{ij}^{obj} (\sqrt{w_i} - \sqrt{\hat{w}_i})^2 + (\sqrt{h_i} - \sqrt{\hat{h}_i})^2 $$

while in forward_detection_layer(), detection_layer.c, different loss calculation:

{% highlight cpp %} *(l.cost) += pow(1-iou, 2); {% endhighlight %}

---

$$

• \sum_{i=0}^{S^2} \sum_{j=0}^{S} \mathcal{1}_{ij}^{obj} (C_i - \hat{C}_i)^2 $$

{% highlight cpp %} *(l.cost) -= l.noobject_scale * pow(l.output[p_index], 2); *(l.cost) += l.object_scale * pow(1-l.output[p_index], 2); {% endhighlight %}

---

$$

• \lambda_{noobj} \sum_{i=0}^{S^2} \sum_{j=0}^{S} \mathcal{1}_{ij}^{obj} (C_i - \hat{C}_i)^2 $$

{% highlight cpp %} (l.cost) += l.noobject_scalepow(l.output[p_index], 2); {% endhighlight %}

---

$$

• \sum_{i=0}^{S^2} \mathcal{1}{i}^{obj} \sum{c \in classes} (p_i(c) - \hat{p}_i(c))^2 $$

{% highlight cpp %} *(l.cost) += l.class_scale * pow(state.truth[truth_index+1+j] - l.output[class_index+j], 2); {% endhighlight %}