Closed JinhangZhu closed 3 years ago
对训练集运用sklearn.cluster.KMeans
计算出来anchors。
37, 53
42, 77
45, 90
54, 96
60, 113
67, 135
76, 136
91, 166
100, 203
在Colab上训练了hand64小数据集,效果非常之差,心情很down,也没想明白什么原因,明天调试一下。
尝试了:
均无改进。
明天尝试以下。为什么?1)找到好的代码版本,一个exploratory process。2)探寻数据集大小的影响。之前的 #2 只用COCO16结果为什么好?思考一下。
[x] Reproduce #8 的左右手训练
[x] 使用全数据集ego-hand训练
计划
[x] 对训练集运用k-means生成anchors
[x] 训练YOLOv3模型
[x] 设计算法接着识别左右手
[ ] SSD review
[ ] Batch normalization