In our approach, we applied a few modi cations to the 50layered Residual Network. Our preliminary experiments with the PlantCLEF 2016 dataset showed that the modi cations improved classi cation performance. We have trained three models based on the modi ed Residual Network con guration with di erent combinations of trusted and noisy PlantCLEF 2017 datasets. Using con dence scores extracted from the three models, we have submitted four runs and our methods showed competitive classi cation performance.
1.1
In ResNet-50, a total of three convolution operations1 with stride size 2 but lter size 1 1 is used for down-sampling. As the lter size is smaller than stride size, part of the activations may be ignored in the ltering i.e. convolution processes, as demonstrated in Figure 1.
Therefore, we reduced stride size of the three convolution operations in ResNet-50 from 2 to 1. In addition to this, max pooling of stride size 2 and lter size 2 2 are inserted before these convolution operations. We label this modi ed con guration as ResNet-50-MP. 1.2 Down-sampling is an essential element in Convolutional Neural Network, which reduces number of activations (as well as computational complexity). However, 1 Namely resf3,4,5ga branch2a based on the ResNet-50 model de nition in github.com/KaimingHe/deep-residual-networks applying down-sampling too early may leave too little activations for subsequent convolution operations, thus impacting classi cation performance.
In a paper by He et al. [
This section illustrates our implementation for the plant identi cation task i.e.
PlantCLEF. We use the Ca e framework by Jia et al. [
Data Augmentation. Training images are randomly scaled such that shorter sides are in the range of [224; 336]. Scaled images are randomly rotated for 45 . Rotated images are randomly cropped into 224 224, and nally the cropped images are randomly horizontal ipped.
As for test images, they are scaled such that the shorter sides become 224. The scaled images are then horizontally ipped. Both ipped and non- ipped images are applied into a trained network for prediction. Class-wise outputs (before softmax normalization) of both instances are averaged and then softmax normalized.
Input Normalization. Mean centering of input is already a common procedure to train or test a network. However, this alone may not be su cient as variance is not considered in this process.
Therefore, we attempted to normalize at higher order by applying Batch
Normalization [
Dataset Preparation. ResNet-50, ResNet-50-MP and ResNet-50-MPD congurations in Section 1 are trained for 100 epochs with PlantCLEF 2016 training dataset. After each training epoch, each con guration is validated with PlantCLEF 2016 test dataset. Omitting unseen images i.e. of ClassId 9999, a total of 113204 training images and 4510 test images are utilized. Note that data augmentation and normalization as detailed in Section 2.1 are applied to all con gurations.
Batch Size. The hardware we use for preliminary experiments is a single NVIDIA's Tesla K40. We use largest possible batch sizes i.e. based on the hardware's memory limitation of 12 GiB.
Learning Schedule. Initial learning rate is 0.1 and is multiplied by 0.1 twice throughout the training process. Training iterations for each learning rate is divided with ratio 4:2:1 across 100 training epochs.
In summary, batch sizes for both original ResNet-50 and ResNet-50-MP are maximized at 31 and they are trained with learning rates 0.1, 0.01, 0.001 for 208671, 104336, 52168 iterations respectively. As for ResNet-50-MPD, the largest possible batch size is 21 and this con guration is trained with the same learning rates for 308038, 154019, 77010 iterations respectively. Validation accuracy of the whole training process for all three con gurations is shown in Figure 2. 2.3
Based on the results as shown in Figure 2, we selected the ResNet-50-MPD con guration for this year's plant identi cation task.
Dataset Preparation. Out of the 256287 trusted training images provided by
the task organizers of PlantCLEF 2017 [
1.0 0.8
Di erent combinations of trusted and noisy images are used to train the ResNet-50-MPD con guration. Our rst model is trained with trusted training images only2, and our second model is trained with noisy training images only3, while our third model is trained with both mixed together. We label the rst as model T, the second as model N, and the third as model X. All three models are validated with the 25063 validation images.
Batch Size. We use a single NVIDIA's Quadro P6000 to train the three models. With 24 GiB memory, we were able to use batch size up to 47. Learning Schedule. As the trusted and noisy datasets are a lot larger compared to last year's, we were not able to train for 100 epochs but a xed amount of training iterations. All three models are trained for 350000 iterations: learning rate 0.1 for 200000 iterations, 0.01 for 100000 iterations and 0.001 for 50000 iterations. In other words, model T was trained for around 71 epochs, model N was trained for around 12 epochs, while model X was trained for around 10 epochs. Validation accuracy of the training processes is summarized in Figure 3. Run Submission. A total of 25170 test images are provided by the task organizers. As detailed Subsection 2.1, each test image is scaled and then horizontally ipped. Both ipped and non- ipped test images are applied into all 2 Corresponds to training set E in imageclef.org/lifeclef/2017/plant 3 Corresponds to training set W 1.0 0.8 three models i.e. model T, N and M. For each model, class-wise average of condence scores (before softmax normalization) extracted from both ipped and non- ipped images is computed. Additionally, class-wise average of con dence scores with the same ObservationId is also computed. The averaged con dence scores are then softmax normalized. We have submitted four runs with team name KDETUT for this year's plant identi cation task, as summarized below. { KDETUT Run 1: Based on model T (trained with trusted images only) { KDETUT Run 2: Based on model N (trained with noisy images only) { KDETUT Run 3: Based on model X (trained with trusted and noisy img.) { KDETUT Run 4: Average of con dence scores based on model T and N 3
Mean Reciprocal Rank (MRR) is used as evaluation metric. Evaluation results released by the task organizers are shown in Table 1 and Figure 4. Among the submitted four runs, Run 4, which is average of con dence scores extracted from model T and N, shows the best classi cation performance.
In this paper, we described our approach to PlantCLEF 2017, focusing on some modi cations to the 50-layered Residual Network. Nevertheless, there are still rooms for improvements in our approaches, as itemized below.
{ Especially models trained with very large datasets i.e. model N and X, the xed amount of 350000 training iterations may not be su cient. For example, as detailed in Subsection 2.3, in fact model X is only trained for around 10 epochs. However, this setup already requires 4 days to train each model. More training iterations i.e. longer training time may be required to achieve superior classi cation performance. { Among the four runs we have submitted, Run 4 yields the highest classication performance. It is based on average of con dence scores computed from model T and N (one each). We believe classi cation performance can be further improved if the con dence scores are averaged from even more models, for example ve model T and ve model N. This is however at the cost of multiplied computation time.