Regularizing Reasons for Outfit Evaluation with Gradient Penalty

02/02/2020 ∙ by Xingxing Zou, et al. ∙ Duke University Hong Kong Polytechnic University The Chinese University of Hong Kong 0

In this paper, we build an outfit evaluation system which provides feedbacks consisting of a judgment with a convincing explanation. The system is trained in a supervised manner which faithfully follows the domain knowledge in fashion. We create the EVALUATION3 dataset which is annotated with judgment, the decisive reason for the judgment, and all corresponding attributes (e.g. print, silhouette, and material .). In the training process, features of all attributes in an outfit are first extracted and then concatenated as the input for the intra-factor compatibility net. Then, the inter-factor compatibility net is used to compute the loss for judgment. We penalize the gradient of judgment loss of so that our Grad-CAM-like reason is regularized to be consistent with the labeled reason. In inference, according to the obtained information of judgment, reason, and attributes, a user-friendly explanation sentence is generated by the pre-defined templates. The experimental results show that the obtained network combines the advantages of high precision and good interpretation.

READ FULL TEXT VIEW PDF
POST COMMENT

Comments

There are no comments yet.

Authors

page 1

page 2

page 3

page 4

This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Introduction

Fashion compatibility evaluation is closely related to our daily life (e.g. Echo Look [10]), and it has attracted increasing attention from researchers [7, 6, 35]. Mainstream methods for fashion compatibility evaluation adopt metric learning: fashion items of the outfit are embedded into a common compatibility space, where items that appear in the dataset are closer in representation and otherwise, have a farther distance. They assume that the occurrence rate of an outfit has direct relevance with its compatibility, which effectively equates the concepts of common and uncommon to compatible and incompatible. However, a discrepancy between being common and being good exists in fashion,. A very common outfit is more likely to be normal rather than good. How to provide professional evaluations that give convincing judgments of good, normal and bad, is still open.

Figure 1: Fashion compatibility evaluation with reason. Our evaluation system offers a clear judgment with convincing explanations.

A few efforts further focused on giving explanations for the output judgments. [19, 5] took user reviews as training data to generate textual explanations. [13] analyzed the outfit images and used heat maps as their explanation. [32] decomposed the item images into human-interpretable features, and identified the most influential feature that contributes to the output. However, the explanation generated by those methods is short on convincingness because of the following limitations: 1. Not relate to specific and concrete reasons. The textual explanation could be very vague, e.g. “This dress is so beautiful, I like it”. 2. Lack of domain expertise. It might simply recognize fashion attributes rather than analyze their relation, e.g. “This orange T-shirt and black pants”. 3. Not aligned with human experience. Heatmaps may attend to image regions that are hard for the human to comprehend.

In this work, we build an outfit evaluation system that faithfully respects the domain knowledge in fashion. The judgment is summarized into three levels: the outfit is good, bad, or normal. For example, the first outfit in Figure 1 looks bad because the unmatched print types between the top and bottom make its appearance too dazzling. The mismatch in print is the logical reason to form the judgment. The normal is a common situation when evaluating outfits. For example the second outfit in Figure 1, an ordinary black T-shirt with black jeans, does not reach the bar of a good mix and match, but not bad either. While it is hard to single out a concrete reason for being normal, we can give an explanation for the judgment of good and bad. In fashion, the major factors for evaluating outfits include color, print, material, etc[8]. The judgment is based on the overall visual expression of those interplaying factors. An outfit is regarded as bad as long as one factor is not well-matched. If all factors arrive at visual harmony, then the outfit at least can be put into normal. Moreover, a good one must possess certain special design to make it stand out from a normal one. In summary, normal is the intermediate level between good and bad; and for good and bad, we further expect a concrete reason for its deviation from normal.

In correspondence to the above discussion, we prepared a dataset that consists of outfit images annotated with the judgment and its decisive reason. To realize the evaluation with reason task, we propose an explainable evaluation framework as follows. Given an outfit composed of several fashion items, feature vectors for color, print, material, silhouette, and design details are first extracted. For each of the factors, an independent net is used to produce an intra-factor compatibility feature. Then all intra-factor compatibility features are concatenated and input into the inter-factor compatibility net, which outputs the judgment. Reason for the judgment is traced back by computing gradients of the judgment w.r.t. the previous intra-factor compatibility features, in a way that resembles Grad-CAM 

[26]

. To increase reasonableness other than simply interpreting as-is, we enforce the traced reason to align with annotated reason by adding a regularization in the form of gradient penalty. Based on the results obtained by the network, a user-friendly explanation is generated based on the pre-designed decision tree.

Our main contributions include: (1) We formulate the fashion compatibility evaluation with reason task in a new framework, which respects the domain knowledge in fashion. (2) We annotate an outfit evaluation dataset EVALUATION3 which has 18,108 pairs of outfit with judgments, reasons and attributes. (3) We use gradient penalty to align the explanation of the network decisions to that of expert’s. Dataset and source code will be released shortly.

2 Related Work

Visual compatibility learning.

Simo-Serra et al[28] learned the fashionability using Conditional Random Field. Oramas et al[23] adopted mid-level attributes to measure the compatibility. Li et al[18] used annotated quality scores to supervise the grading of outfits. Han et al[11] used Bi-LSTM, and Visileva et al[33] improved on it by considering type information. Recent works built upon the clothing embedding [15, 35]

, which can be learned through autoencoder or by the Bayesian Personalized Ranking (BPR) loss 

[30], the hinge loss [16], the triple loss [4], and the binary cross-entropy loss [27]. Another line of work models the clothing style [20, 34, 14, 1], which expresses the compatibility implicitly. In a word, mainstream approaches mainly adopt the relative embeddings to compute evaluation scores. The benefit is not needing explicitly graded data for they assume all outfits in the dataset are positive samples. However, it is unreasonable to equate the observation of high occurrence rate with the concept of good. Meanwhile, a relative score is less friendly to users and the compatibility space is hard to interpret. Our work differs from above in several aspects. Firstly, we adopt the fashion-matching principles [8] as the standard for evaluating whether an outfit is good or bad. Secondly, we propose absolute ratings with three levels as judgments in the evaluation scene. Thirdly, we are interested in the explainability of the given judgment.

Explainable outfit recommendation.

Yang et al[36] introduced an attribute-based interpretable compatibility method. They mixed all attributes and found the informative attribute crosses statistically, whereas we find the dominant factor by mimicking the analyzing process of fashion experts. Feng et al[9] proposed a partitioned embedding network, where color, shape and texture are defined as the main factors, and the score of each factor is used as the explanation. Lin et al[19] presented a recommendation system and used the generated comments as explanations. Xu et al[5] adopted user reviews into an attention-based architecture to enhance the performance of recommendation and the interpretability. These methods provide reasons using the heatmap, comments or reviews, in contrast, our explanation is inferred based on the trained network. Recently, Tangseng et al[32] defined an influence score of factors using gradients, which has a similar idea to this work. The major difference is that, they analyzed the trained model post-hoc, while we explicitly supervise the explanation of the model to be aligned with expert interpretations.

Explaining DNN classifiers.

Works on explaining neural networks greatly improve the interpretability of the black-box deep models. Springenberg

et al[31] visualized CNN predictions by highlighting contributing pixels. Zhou et al[37] proposed the Class Activation Mapping (CAM) for visualization. Selvaraju et al[26] presented Grad-CAM which adopted the gradients of any target concept to enhance localization. Chattopadhyay et al[3] further improved the performance with Grad-CAM++. Regularizations can also be added [2, 25, 21] to improve the explanation quality. However, directly using the heatmap produced by these methods as the reason for fashion judgments is insufficient, because the reason for the judgment is not a local region of an item image but the global factors such as color, print, and silhouette. To overcome this limitation, we trace the reason for the judgment back to human-interpretable factors, rather than pixel-wise heatmaps. Traditional methods like decision tree also own limited explainablility [17]

. They are not suitable for our task because 1. For any outfit, tree based methods give a decision path from root to leaf, which is hard to locate a single, dominate factor that contributes to the decision. 2. In the tree construction, the space is split based on heuristics such as information gain which is not directly relevant.

3 Dataset Construction

Based on the perspective of fashion, we grade an outfit into three progressive levels: bad, normal, and good. The bad level is defined as the outfit having something wrong, e.g. a wrong color matching or a dazzling print. If the outfit does not make any mistake that breaks the visual balance, it comes up to the normal level. Further, if the outfit has a special design, e.g. attractive color matching, special print, or good cutting, it reaches the good level. Note the logical connection among the three levels. If the outfit has something wrong, i.e. some factors not well-matched, it will be regarded as bad, whether it has some special design or not.

Our evaluation system needs three-level judgment with the corresponding reason. Detailed attribute annotations are also needed to train factor feature extractors. However, existing public fashion datasets do not have labeled judgments and their decisive reasons. Thus we introduce a new dataset, namely EVALUATION3. The image source is a subset of the Polyvore dataset [11]. The dataset contains four sets including train set, validation set, test set, and test-random set. All labels are manually annotated by 9 independent annotators (all major in fashion). Voting mechanism was adopted to get the final label. Furthermore, to mitigate the influence of cultural backgrounds, the 9 annotators (4 male and 5 female) were selected from different fashion regions including Asia-Pacific, Europe, North America, etc.

Figure 2: Samples of labeled attributes in the EVALUATION3, which cover material, silhouette, and design details. We also show the five main colors, and their histograms.

Each outfit contains a top image and a bottom image. We annotated 18,108 outfits in total, including 2,861 (15.8%) in good, 13,587 (75.0%) in normal, and 1,660 (9.2%) in bad. Among them, 12,608 outfits (15.7% in good, 75.4% in normal, and 8.8% in bad) belong to train set, 2,500 outfits (17.0% in good, 74.4% in normal, and 8.6% in bad) are put into validation set, and the rest 3,000 outfits (15.6% in good, 73.8% in normal, and 10.6% in bad) are used for testing. In particular, we added a test-random set that randomly pairs tops and bottoms in the test set. It contains 3,000 outfits with 22.8% in good, 61.2% in normal, and 15.4% in bad. This set provides a different data distribution because outfits in the original dataset are all paired by Internet users.

As our outfit evaluation is based on features of the design factors, the accuracy of attribute recognition directly affects the evaluation result. To reduce domain gap, we labeled the corresponding attributes on our dataset instead of adopting different fashion attribute datasets. We summarized 14 print types, including abstract, allover, animal print, etc. Besides, there are 5, 8 and 10 attributes in silhouettes (A-line, H-line, Peg-top, etc.), design details (tiered, wrap, ruffle etc.), and materials (knit, lace, leather, etc.), respectively. Figure 2 shows some examples of the dataset.

4 Approach

Given an outfit that comprises of a top and a bottom, we first classify it into one of the three levels:

good, normal, and bad. Except for normal, a reason will then be identified among three choices: color, print, and design. Here design is a collective name for material, silhouette, and design details. The obtained results, including judgment, reason, and attributes, together generate the final explanation based on a pre-designed decision tree. For convenience, we denote the judgment set by = {good, normal, bad}, the reason set by = {color, print, design}.

4.1 Architecture

The pipeline of computing judgment and reason shown in Figure 3

. The judgment is obtained in three stages, namely (1) factor feature representation, (2) intra-factor compatibility and (3) inter-factor compatibility. For each pair of outfit images, we attain the feature representation via the respective function or convolutional neural networks. Then the features of each factor first pass through the intra-factor compatible net, and produce the intra-factor compatible feature which mixes the up and bottom information. At last, we use these intra-factor compatible features for the final judgment. The contribution of factors to the decision is traced back via gradients to the intra-factor compatible feature, and we can calculate the prominent reason for the judgment. To generate user-friendly explanations, we added a post-processing step using a sentence template, which is completed by the elements from a pre-designed decision tree. Note that our evaluation framework can be extended to process the varied number of fashion items simply by replacing the CNNs in stage 2 to LSTMs 

[11]. In the following, we will discuss the network design, the process of tracing back reasons and the way of explanation generation in detail.

Figure 3:

The pipeline of outfit evaluation network. The features of each contributing factor are extracted by the feature extraction net. Then, the judgment and reasons are trained in the same network.

Fashion feature extraction.

For each input image, we have five feature vectors based on the five fashion compatibility evaluation factors, i.e

. color, print, material, silhouette, and design details. Features of the print, material, silhouette and design details are extracted by deep models. We fine-tune the ImageNet pre-trained ResNet-18 

[12] on the EVALUATION3. The last feature map (

-dimensional) is used as the representation of these fashion factors. Take the print feature net for example. We 1. change the output neurons of the last fc layer of ResNet from 1,000 to 14; 2. initialize parameters using pre-trained ImageNet models; 3. train the last fc layer on the dataset for 10 epochs with other parameters fixed; 4. jointly train all parameters for 30 epochs. For the color feature, we compute the color histogram, The

Fashion Color system (FOCO) [38] is used to set up bins of the histogram. FOCO divides the H, S, B channels into 15, 8 and 6 levels, respectively. For example, PANTONE [24] Candy Apple Red is represented as (1, 8, 4) in FOCO. We concatenate the five major colors with their ratio and obtain a -dimensional color feature.

Intra-factor compatibility network.

For each outfit and each factor, we obtain two feature vectors corresponding to the top and the bottom. Then, the two vectors are fused by an intra-factor compatibility network, which is a three-layer fully-connected network. The output feature represents the interaction between the top and the bottom on a specific factor. In the end, all five output features are concatenated as a single intra-factor compatibility feature vector and sent to the next stage. This intra-factor compatibility feature is also the place where we trace the reason back to.

Inter-factor compatibility network.

This network (also has three fully-connected layers) captures how all factors are inter-related to the final decision. The input is the five intra-factor compatibility features, and the output is the probability of being judged as

good, normal or bad. The cross-entropy loss for judgments is computed therefrom.

4.2 Reason for Judgments

Grad-CAM [26]

is widely used for inspecting the spatial contributing regions of inputs. Conditioned on a class label, it finds the dominant neurons contributing to the decision toward this label by tracking back the gradient of the label logit (

i.e. logit before softmax) w.r.t. feature maps. For any class label , the gradient of the logit with respect to the -th feature map of a convolutional layer, is computed. Then these gradients are global-average-pooled to obtain the neuron importance weights ,

(1)

where iterates over the spatial dimensions and is the number of pixels in the feature map. The product of the neuron importance weights and the forward activation map is obtained as the heatmap ,

(2)

Different from Grad-CAM, we care about the “heatmap” over human-interpretable factors rather than images pixels. Denote the logit for the judgment by . Let the intra-factor compatibility feature be with elements . We define the contribution of each element for the decision of judgment as ,

(3)

As shown in Figure 3, the intra-factor compatible feature is divided into five distinct parts, each represents the compatible situations of one factor, such as color or print. Suppose the index set of neurons for factor is , then we define the contribution of the factor as the average of the contribution of the constituent elements. and the positive contribution of this factor for the judgment as

(4)

The positive contribution in Equation (4) is slightly different from the form of Grad-CAM, in which the is put inside the summation. Unlike Grad-CAM, whose goal is to analyze the network after training, we actively regularize the network to be explainable during the training. Formulation of Equation (3) and (4) is based on the following contribution assumption; we also introduce the relativeness assumption.

(1) Contribution assumption: the contribution of a neuron for the decision consists of two parts: the saliency and the sensitivity. Firstly the value should be present, and the impact is greater if the value is larger. If it is negative, then it will be filtered out immediately by the subsequent activation. Secondly, the decision should be sensitive to the change of the neuron, which is measured by the derivatives [29]. If the value is large but the derivative is zero, it means that the decision is irrelevant to the presence of the neuron. The product of saliency and sensitivity has been used previously, such as in the Grad-CAM, the Item Feature Influence Value (IFIV) [32], and the relevance score in [22] which is derived from the perspective of Taylor decomposition.

(2) Relativeness assumption: the difference between two contributions has semantic meaning. This motivates from the fashion domain knowledge: both the good and bad levels are defined relative to the normal level. What people care the most for the good is its highlight comparing to a normal one, rather than a full list of the contributing factors. Good color matching, nice print combination, etc., all may contribute to its being good but not necessarily stand out. The same argument applies to the bad level. We aim to find the factor which contributes the most to good compared with normal. Applying the relativeness assumption specifically to the evaluation task at hand, the mathematical form of the reason for the good judgment can be formulated as

(5)

We use the good minus normal to estimate to which extent will each factor lead to good rather than normal, and use the

operator to extract the dominant factor. Similarly, bad is bad because it is worse than normal. The reason is

(6)

For normal cases, there is no explicit reason available because all factors are neither outstanding nor terrible.

4.3 Supervise Reason with Gradient Penalty

We want the prominent reason output by the network to be aligned with pre-labeled data. This is achieved by training the network with specially designed regularizations. We propose three forms of regularizations and compare their performances in the experiment section. Let be the traced reason vector for the ground-truth class, where

(7)

Here , is an indicator function for ground-truth judgment . When judgment is the same as ground-truth, ; otherwise, . e.g. if represents good. The three components of are the reason strength of factor color, print and design, respectively. Denote the ground-truth reason as , then the three regularizations can be expressed as the following. The reason shown in the formulation below refers to the label of reason, i.e. color, print and design, respectively.

  1. Cross-entropy (CE) regularizer,

    (8)
  2. Linear regularizer. If the prediction is wrong, then linearly pull up the ground-truth reason strength and push down the wrong one,

    (9)
  3. Square regularizer,

    (10)

The total loss is the combination of the judgment loss and the reason loss, i.e. one of the above regularizers,

(11)

where is a hyper-parameter that controls the effect of reason regularization. The and is the loss of the judgment and reason, respectively. Since the gradient appears in the definition of contribution and reason (Equations (4) and (7)), the loss term penalizes the gradient. The gradient penalty directly affects the network parameters after the compatible feature layer. In the definition, feature map also presents. This means that the parameters of the intra-factor compatible net, those before the compatible feature layer, are also regularized.

4.4 Explanation Generation

Figure 4: The pipeline of explanation generation with the obtained judgment, reason and attributes of an outfit.

To obtain explanations that are friendly to users, we used an explanation template in the form of the decision tree (manually designed, not learnable, see Figure 4

), which is built with the assist of fashion experts. This part can also be implemented using different approaches such as text generation techniques. The template is sufficient for our demonstration purpose. After analyzing an outfit as described earlier, we obtain the results including judgment, its decisive reason, and all corresponding attributes of the outfit. Then, the explanation is generated accordingly following the decision tree. For example, as shown in Figure 

4, the judgment is bad while the predicted reason is print. From the prescribed sentence template: “The print_t top and the print_b bottom make the outfit too dazzling”, we can obtain the explanation: “This outfit is bad. The floral print top and the floral bottom make the outfit too dazzling.” We found that it is unnecessary to mentioned attributes other than the main factor, e.g. the short-length of the skirt, because they have little effect on whether the outfit is compatible.

judge reason partition explain test-acc-j test-acc-r test-r-acc-j test-r-acc-r
Multi-CLS-Part
IFIV [32]
Reason-NoReg
Ours linear
Ours square
Ours cross-entropy
Table 1: Comparing our method with related methods. All experiments were repeated 5 times. The first line is multi-task (judgment and reason) classification models, where Multi-CLS-Part uses the same fashion-related features as our model. Both IFIV and Reason-NoReg did not take the supervision of reason. For IFIV, we did not implement temperature-scaling, since only the factor with maximum contribution was considered. The rest are our method with three different regularizations.

5 Experiment

Figure 5: The reason and contribution factor analysis on some outfits via a model with cross-entropy reason-regularization. At the top of each mini-table, good/good means ground-truth/predicted judgment, and the ground-truth reason is also shown. In each mini-table, the three columns represent the component of factors color, print and design, respectively. Rows with C are the contribution (see Equation (3)) of predicted judgment, G are the reason for good (see Equation (5)) and B are reason for bad (see Equation (6)).
Figure 6: The samples of our fashion compatibility evaluation including judgment, reason and explanation.

In this section, we first conduct experiments on EVALUATION3 to show the advantages of the proposed evaluation system quantitatively and qualitatively. The quantitative analysis focuses on both the judgment accuracy and the reason accuracy. We show the superiority of our approach on the test set and the test-random set. Recall that the test-random set is the random composition of fashion items in the test set, which could be regarded as an indicator of the model’s generalization ability. Regarding qualitative analysis, we visualize the detailed contributions in varied situations, e.g. judgment: bad, reason: print, to verify the advantage of our method with gradient penalty. Then, we provide ablation study on the choice of contribution formulations and regularizers.

5.1 Compatibility Evaluation with Reasons

Metric.

We measure both the judgment accuracy and the reason accuracy. The judgment accuracy is computed as the number of correct predicted judgment (compared with the labeled ground-truth of judgment) divided by the number of total predicted samples. It is worthwhile to note that, for an outfit, if the judgment is wrong, then the reason for this judgment may not make sense. Thus, the premise of the correctly predicted reason is that the judgment is correct. Under the condition of the right predicted judgment, we count the number of correctly predicted reason. Then divided it by the number of total predicted outfits. Then divide it by the number of correct predicted judgments.

Training details.

We train the pipeline (Figure 3) in two steps. In the first step, we train four CNNs to learn the representations of the print, the material, the silhouette and the design details separately. The CNN structure is ResNet18. Each factor has its corresponding labels, with examples shown in Figure 2

. All the networks are trained using RMSProp with initial learning rate

and weight decay

, the learning rate is divided by 10 every 30 epochs. Parameters of the pre-trained model are fixed for the first 20 epochs. We use the feature of the second last fc-layer from the pre-trained model as the input to the second stage. The color feature is directly extracted by the histogram. In the second step, we train the intra-factor compatibility networks, the inter-factor compatibility network, and the reason regularization (Modern DL frameworks such as PyTorch have built-in support for computing the gradient of gradients) jointly. In particular, we use SGD with initial learning rate

, and weight decay for epochs and the learning rate is also divided by every 30 epochs. Since the dataset is unbalanced (the ratio of normal in the train set reaches 75.4%), we adjust the sampling rate for each class (good, normal, and bad) to obtain the balanced sampled data.

Quantitative analysis.

Table 1 shows the comparison on the accuracy with related methods. IFIV [32, Equation 10] computes the item feature influence value, which has the same maximum factor as our . We also test the performance without regularization (i.e., set ) as shown in Reason-NoReg. Meanwhile, we evaluate the performance with three regularizations, linear, square, and cross-entropy, respectively. We can see that (1) High judgment accuracy: Our method reaches a high classification accuracy on evaluation without increasing parameters. As shown in the first line in Table 1, the judgment accuracy and reason accuracy of the test set are and . These two accuracies on the test-random set are and . We can see that our result even surpasses to pure-classification models w.r.t. regarding the judgment and reason as two independent classification problems and jointly train via multitasking. Meanwhile, under the setting of cross-entropy loss, the judgment accuracy is larger than Reason-NoReg which has no regularization. This demonstrates the effectiveness of the gradient penalty. (2) High reason accuracy: For methods like IFIV and Reason-NoReg, they can bring us some explanations or feedback. However, their results are less likely to be fully aligned to expert thoughts. (3) Good generalization: From the good performance on test-random sets, we see that our result still achieves the highest accuracy among the compared methods, which means our approach enjoys good generalizability.

As discussed in the Section 2, the embedding methods are mainly adopted in previous works. We conducted a study to show that embedding may fail to distinguish good, normal, and bad, which makes it not suitable for our task. There are two basic ways to represent evaluation: absolute rating and relative embedding. Embedding methods assume that there is a common compatibility space for fashion items. The idea is to pull the representations of positive pairs closer and push the negative pairs farther (positive pairs are those that appear in dataset e.g. Polyvore [11] and FashionVC [30], and negative pairs are random combinations). We built an embedding model consisting of two ResNet18 networks which respectively embeds top and bottom to the compatibility space. The compatibility is measured by the cosine distance, and the model is trained using the BPR loss. After training, the average distance between top and bottom of bad is 0.464, far less than that of normal (0.562) and good (0.574). This result is against with the assumption that more compatible items have a smaller distance.

Qualitative analysis.

We show some examples with more details in Figure 5. Take the first mini-table as an example. The ground-truth and predicted judgment for this outfit are both good while the ground-truth reason is color. The scores computed by Equation (7) (the second row with label G) when set = good are 0.647 in color, 0.098 in print, and 0.045 in attribute, which means color is the predicted reason. Similarly, the scores calculated by Equation (3) (the first row with label C) are shown. The red number means the predicted reason is wrong while green one means correct. The first two rows are examples of good, and bad, respectively. We can see that our model provides more accurate judgment with the operation of taking difference. Further, to demonstrate the model indeed learned the concept of color and print, we change the color or print of a good outfit in the last row. The prediction changes accordingly. Furthermore, the reason reflects the changing factor. This demonstrates our model indeed learned the reason for its judgment.

Additionally, we present the final performance of our outfit evaluation system in Figure 6. Given an outfit, an absolute judgment with convincing explanation can be provided to users. Take the outfit at the fifth line in Figure 6 for example. This outfit is bad, because the pink yarrow color in top and cadmium green color in the bottom is wrong color matching. Meanwhile, our method could also be applied in outfit recommendation scene by simply scaling the logit or the probability of the judgment as of the relative score for recommendation sort. This kind of explanation framework took the suggestions of fashion experts as reference.

Figure 7: The reason validation accuracy for different formulations in the unsupervised setting, i.e. without any reason-related regularizations. All experiments are repeated times. Here the notation “good” represents and “” is shorthand for , etc. We can see that the 6-th formulation positive contribution difference has the greatest accuracy on average.
Figure 8: The change of judgment accuracies and reason accuracies with different regularization coefficient . We show three regularization methods: cross-entropy, linear and square. All experiments were repeated times, and we draw the average validation accuracies on the figure.

5.2 Ablation study

Contribution formulations.

We explore other candidate formulations for the reasons in fashion compatibility evaluation. The network in Figure 3 was trained without any reason regularizations, i.e. using an unsupervised treatment for reasons. After the training, we analyze the reason accuracy for different reason formulations in Figure 7. The first one is to directly use the contribution defined in Equation (3

) and we use it as the reference performance. Especially, Grad-CAM is equivalent to the first one when we take the maximum for the major reason. We see that the contribution difference (the 2nd formulation) and the positive contribution difference (the 6th formulation) are slightly better than the baseline. Particularly, the positive contribution difference has lower variance and thus we adopt the 6th one as the formulation for reason.

Regularizers.

Meanwhile, we conduct experiments on three different forms of regularization: (a) cross-entropy, (b) linear, and (c) square. The judgment accuracies and the reason accuracies of these formulations are plotted in Figure 8, in which the regularization coefficient varies. For all three regularization methods, the trade-off between judgment accuracy and the reason accuracy. When increases, the reason accuracy generally improves. The judgment accuracy can maintain at the same level for some time, and finally drops when becomes too large. The cross-entropy regularization is most sensitive to the change of hyper-parameter, and reaches the best judgment & reason accuracy trade-off at . The best value for linear and square formulations are and , respectively.

6 Conclusion

In this work, we present an outfit evaluation system with the feedback consisting of an absolute judgment together with an explanation. In particular, the interpretation of neural networks produced by our method is aligned with expert annotated reasons by adding a gradient penalty regularization. We use the difference of contributions to model the interaction of the two judgments of good and bad with the normal judgment. With the reason regularization term imposed on the loss, we can make the network more interpretable without losing its performance. Comprehensive experiments conducted on the newly annotated EVALUATION3 show the effectiveness of the proposed method.

References

  • [1] Z. Al-Halah, R. Stiefelhagen, and K. Grauman (2017) Fashion forward: forecasting visual style in fashion. In

    Proceedings of the IEEE International Conference on Computer Vision

    ,
    pp. 388–397. Cited by: §2.
  • [2] M. Al-Shedivat, A. Dubey, and E. P. Xing (2017) Contextual explanation networks. arXiv preprint arXiv:1705.10301. Cited by: §2.
  • [3] A. Chattopadhay, A. Sarkar, P. Howlader, and V. N. Balasubramanian (2018) Grad-cam++: generalized gradient-based visual explanations for deep convolutional networks. In 2018 IEEE Winter Conference on Applications of Computer Vision, pp. 839–847. Cited by: §2.
  • [4] L. Chen and Y. He (2018) Dress fashionably: learn fashion collocation with deep mixed-category metric learning. In

    Thirty-Second AAAI Conference on Artificial Intelligence

    ,
    Cited by: §2.
  • [5] X. Chen, Y. Zhang, H. Xu, Y. Cao, Z. Qin, and H. Zha (2018) Visually explainable recommendation. preprint arXiv:1801.10288. Cited by: §1, §2.
  • [6] G. Cucurull, P. Taslakian, and D. Vazquez (2019-06) Context-aware visual compatibility prediction. In

    The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    ,
    Cited by: §1.
  • [7] Z. Cui, Z. Li, S. Wu, X. Zhang, and L. Wang (2019) Dressing as a whole: outfit compatibility learning based on node-wise graph neural networks. arXiv preprint arXiv:1902.08009. Cited by: §1.
  • [8] M. Eckman and J. Wagner (1995) Aesthetic aspects of the consumption of fashion design: the conceptual and empirical challenge. ACR North American Advances. Cited by: §1, §2.
  • [9] Z. Feng, Z. Yu, Y. Yang, Y. Jing, J. Jiang, and M. Song (2018) Interpretable partitioned embedding for customized multi-item fashion outfit composition. In Proceedings of the 2018 ACM on International Conference on Multimedia Retrieval, pp. 143–151. Cited by: §2.
  • [10] S. Gibbs (2017) Amazon unveils echo look, a selfie camera to help you choose what to wear. Note: https://www.theguardian.com/technology/2017/apr/26/amazon-echo-look-webcam-choose-fashion-outfits-alexa-smart-selfie-camera Cited by: §1.
  • [11] X. Han, Z. Wu, Y. Jiang, and L. S. Davis (2017) Learning fashion compatibility with bidirectional lstms. In Proceedings of the 2017 ACM on Multimedia Conference, pp. 1078–1086. Cited by: §2, §3, §4.1, §5.1.
  • [12] K. He, X. Zhang, S. Ren, and J. Sun (2016) Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 770–778. Cited by: §4.1.
  • [13] M. Hou, L. Wu, E. Chen, Z. Li, V. W. Zheng, and Q. Liu (2019) Explainable fashion recommendation: a semantic attribute region guided approach. Twenty-Eight International Joint Conference on Artificial Intelligence. Cited by: §1.
  • [14] W. Hsiao and K. Grauman (2017) Learning the latent “look”: unsupervised discovery of a style-coherent embedding from fashion images. In 2017 IEEE International Conference on Computer Vision (ICCV), pp. 4213–4222. Cited by: §2.
  • [15] W. Hsiao and K. Grauman (2018) Creating capsule wardrobes from fashion images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 7161–7170. Cited by: §2.
  • [16] W. Kang, E. Kim, J. Leskovec, C. Rosenberg, and J. McAuley (2018) Complete the look: scene-based complementary product recommendation. arXiv preprint arXiv:1812.01748. Cited by: §2.
  • [17] B. Kim and F. Doshi-Velez (2018)

    Introduction to interpretable machine learning

    .
    Proceedings of the CVPR Tutorial on Interpretable Machine Learning for Computer Vision. https://interpretablevision. github. io/index_cvpr2018. html. Cited by: §2.
  • [18] Y. Li, L. Cao, J. Zhu, and J. Luo (2017)

    Mining fashion outfit composition using an end-to-end deep learning approach on set data

    .
    IEEE Transactions on Multimedia 19 (8), pp. 1946–1955. Cited by: §2.
  • [19] Y. Lin, P. Ren, Z. Chen, Z. Ren, J. Ma, and M. de Rijke (2018) Explainable fashion recommendation with joint outfit matching and comment generation. arXiv preprint arXiv:1806.08977. Cited by: §1, §2.
  • [20] J. McAuley, C. Targett, Q. Shi, and A. Van Den Hengel (2015) Image-based recommendations on styles and substitutes. In Proceedings of the 38th International ACM SIGIR Conference on Research and Development in Information Retrieval, pp. 43–52. Cited by: §2.
  • [21] D. A. Melis and T. Jaakkola (2018) Towards robust interpretability with self-explaining neural networks. In Advances in Neural Information Processing Systems, pp. 7786–7795. Cited by: §2.
  • [22] G. Montavon, W. Samek, and K. Müller (2018) Methods for interpreting and understanding deep neural networks. Digital Signal Processing 73, pp. 1–15. Cited by: §4.2.
  • [23] J. Oramas and T. Tuytelaars (2016) Modeling visual compatibility through hierarchical mid-level elements. arXiv preprint arXiv:1604.00036. Cited by: §2.
  • [24] Pantone (2019) Pantone color, chips & color guides — color inspiration. Note: https://www.pantone.com/ Cited by: §4.1.
  • [25] G. Plumb, M. Al-Shedivat, E. Xing, and A. Talwalkar (2019) Regularizing black-box models for improved interpretability. arXiv preprint arXiv:1902.06787. Cited by: §2.
  • [26] R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra (2017) Grad-cam: visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE International Conference on Computer Vision, pp. 618–626. Cited by: §1, §2, §4.2.
  • [27] Y. Shih, K. Chang, H. Lin, and M. Sun (2018) Compatibility family learning for item recommendation and generation. In Thirty-Second AAAI Conference on Artificial Intelligence, Cited by: §2.
  • [28] E. Simo-Serra, S. Fidler, F. Moreno-Noguer, and R. Urtasun (2015) Neuroaesthetics in fashion: modeling the perception of fashionability. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 869–877. Cited by: §2.
  • [29] K. Simonyan, A. Vedaldi, and A. Zisserman (2013) Deep inside convolutional networks: visualising image classification models and saliency maps. arXiv preprint arXiv:1312.6034. Cited by: §4.2.
  • [30] X. Song, F. Feng, J. Liu, Z. Li, L. Nie, and J. Ma (2017) Neurostylist: neural compatibility modeling for clothing matching. In Proceedings of the 2017 ACM on Multimedia Conference, pp. 753–761. Cited by: §2, §5.1.
  • [31] J. T. Springenberg, A. Dosovitskiy, T. Brox, and M. Riedmiller (2014) Striving for simplicity: the all convolutional net. arXiv preprint arXiv:1412.6806. Cited by: §2.
  • [32] P. Tangseng and T. Okatani (2019) Toward explainable fashion recommendation. arXiv preprint arXiv:1901.04870. Cited by: §1, §2, §4.2, Table 1, §5.1.
  • [33] M. I. Vasileva, B. A. Plummer, K. Dusad, S. Rajpal, R. Kumar, and D. Forsyth (2018) Learning type-aware embeddings for fashion compatibility. In Proceedings of the European Conference on Computer Vision (ECCV), pp. 390–405. Cited by: §2.
  • [34] A. Veit, B. Kovacs, S. Bell, J. McAuley, K. Bala, and S. Belongie (2015) Learning visual clothing style with heterogeneous dyadic co-occurrences. In Proceedings of the IEEE International Conference on Computer Vision, pp. 4642–4650. Cited by: §2.
  • [35] X. Wang, B. Wu, Y. Ye, and Y. Zhong (2019) Outfit compatibility prediction and diagnosis with multi-layered comparison network. In Proceedings of the 2019 ACM on Multimedia Conference, Cited by: §1, §2.
  • [36] X. Yang, X. He, X. Wang, Y. Ma, F. Feng, M. Wang, and T. Chua (2019) Interpretable fashion matching with rich attributes. Cited by: §2.
  • [37] B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba (2016)

    Learning deep features for discriminative localization

    .
    In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 2921–2929. Cited by: §2.
  • [38] X. Zou, W. K. Wong, C. Gao, and J. Zhou (2019) FoCo system: a tool to bridge the domain gap between fashion and artificial intelligence. International Journal of Clothing Science and Technology. Cited by: §4.1.