-
Notifications
You must be signed in to change notification settings - Fork 53
Home
实验过程中,发现PCA使用不同数据源对CDW的影响是比较大的。下面是PCA采用不同数据源对CDW的影响实验,实验在oxford上进行,评鉴准则采用mAP。
实验设置:crop、no query expansion、do PCA
数据源 | 维度 | mAP |
---|---|---|
oxford | 512 | 60.2155% |
oxford | 256 | 64.3746% |
oxford | 128 | 66.9665% |
oxford | 64 | 64.3458% |
oxford | 32 | 58.0331% |
数据源 | 维度 | mAP |
---|---|---|
paris | 512 | 70.8359% |
paris | 256 | 69.6122% |
paris | 128 | 64.0718% |
paris | 64 | 58.4009% |
paris | 32 | 52.5268% |
mAP在小数点后又微小浮动。
实验设置:no crop、no query expansion, do PCA
数据源 | 维度 | mAP |
---|---|---|
oxford | 512 | 59.9517% |
oxford | 256 | 64.3746% |
oxford | 128 | 66.9665% |
oxford | 64 | 64.3458% |
oxford | 32 | 58.0331% |
实验设置:no crop、query expansion、do PCA
top@K | 维度 | mAP |
---|---|---|
0 | 512 | 59.9517% |
1 | 512 | 59.9517% |
2 | 512 | 63.7079% |
3 | 512 | 65.6768% |
4 | 512 | 66.7678% |
5 | 512 | 67.4205% |
6 | 512 | 68.3001% |
7 | 512 | 68.9647% |
8 | 512 | 69.5633% |
9 | 512 | 69.5831% |
10 | 512 | 69.8873% |
重构后的代码完成的功能如下:
- 全图提取特征 or 区域框选提取特征
- do PCA or not
- do query expansion or not
- 特征可视化
重构后的检索精度指标评价:do crop, do qe(top@10), do PCA
维度 | mAP |
---|---|
512 | 71.88% |
256 | 72.04%(73.01%) |
128 | 70.33% |
64 | 65.6768% |
32 | 58.8% |
在几百万数据集上获取PCA,然后用在oxford上:
维度 | mAP |
---|---|
128 | 59.15% |
可能的原因:Oxford查询图片都是地标图像集,而这几百万数据集都是短视频中的一些数据,导致获得的主成分不利于地标数据的表达,所以精度降低。 |
do crop, do qe(top@10), do PCA,最高维度256。
维度 | mAP |
---|---|
256 | 62.41% |
We chose the HybridNet for several reasons: first, its ar- chitecture is the same as the famous AlexNet [19]; second, the HybridNet has been trained on the ImageNet subset used for ILSVRC competitions (as many others) and the Places Database [29]; last, but not least, experiments conducted on various datasets demonstrate the good transferability of the learning [29, 12, 9]. Originally proposed in [29], Hybrid- Net has been used in [29, 12, 9]. The results reported in [12] show that deep features extracted from the HybridNet outperforms various architectures trained only on ImageNet, on both InriaHolidays and OxforBuilding benchmarks.