perlex with coord mask & conf_mask

experiencor / keras-yolo2

Easy training on custom dataset. Various backends (MobileNet and SqueezeNet) supported. A YOLO demo to detect raccoon run entirely in brower is accessible at https://git.io/vF7vI (not on Windows).

MIT License

1.73k stars 784 forks source link

perlex with coord mask & conf_mask #453

Open speeding-motor opened 4 years ago

speeding-motor commented 4 years ago

conf_mask why not use y_true mask: if y_true[i,4]=1 , it means there have a box, and conf_mask[i]=1,

why still need to compute the IOU between y_pred with y_true box , and then judge the conf_mask through :conf_mask= conf_mask + tf.to_float(best_ious < 0.6) * (1 - y_true[..., 4])

speeding-motor commented 4 years ago

how to understand the conf_mak here, loss_conf = tf.reduce_sum(tf.square(true_box_conf - pred_box_conf) * conf_mask)

can you please help me ...thank you very much