Recognizing Part Attributes with Insufficient Data

TL;DR

This paper introduces a Concept Sharing Network (CSN) that effectively recognizes part attributes with limited or no training data by disentangling location and appearance features, reducing reliance on extensive annotations.

Contribution

The paper proposes a novel CSN model that enables recognition of part attributes with scarce data and zero-shot capabilities, avoiding the need for detailed part annotations.

Findings

CSN outperforms existing methods on multiple datasets.

Effective recognition of attributes with few training samples.

Zero-shot attribute recognition demonstrated.

Abstract

Recognizing attributes of objects and their parts is important to many computer vision applications. Although great progress has been made to apply object-level recognition, recognizing the attributes of parts remains less applicable since the training data for part attributes recognition is usually scarce especially for internet-scale applications. Furthermore, most existing part attribute recognition methods rely on the part annotation which is more expensive to obtain. To solve the data insufficiency problem and get rid of dependence on the part annotation, we introduce a novel Concept Sharing Network (CSN) for part attribute recognition. A great advantage of CSN is its capability of recognizing the part attribute (a combination of part location and appearance pattern) that has insufficient or zero training data, by learning the part location and appearance pattern respectively from the training data that usually mix them in a single label. Extensive experiments on CUB-200-2011 [51], CelebA [35] and a newly proposed human attribute dataset demonstrate the effectiveness of CSN and its advantages over other methods, especially for the attributes with few training samples. Further experiments show that CSN can also perform zero-shot part attribute recognition. The code will be made available at https://github.com/Zhaoxiangyun/Concept-Sharing-Network.

Tables

Table 1. Table 1: Average precision given 500 training images

attribute	no share	part share	pattern+part share
belly-solid	83.6%	85.9%	85.5%
breast-white	81.3%	82.1%	81.9%
bill-grey	44.7%	46.0%	47.0%
bill-black	78.2%	79.4%	77.0%

Table 2. Table 3: Average Precision(AP) comparison. The numbers are shown only for the 32 attributes that contain more than 1000 positive samples in this dataset. If we perform such comparison for attributes with smaller training set, the baseline always produces very poor result.

Baseline	CAM [69]	STN [21]	PANDA [64]	R-CNN [62]	CSN
63.1%	63.2%	63.7%	64.9%	65.5%*	65.2%

Table 3. Table 4: Average Precision(AP) comparison. The numbers are shown only for the 32 attributes that contain more than 1000 positive samples in this dataset. If we perform such comparison for attributes with smaller training set, the baseline always produces very poor result.

CSN w/o share	CSN	CSN-soft	CSN-soft-1
63.9%	65.2%	65.1%	63.8%

Table 4. Table 5: Average Precision(AP) comparison with different sharing attribute number. N 1 subscript 𝑁 1 N_{1} is the number of sharing part attributes, and N 2 subscript 𝑁 2 N_{2} is the number of sharing pattern attributes. Note: curved attribute only exists in bill-curved , so (d) for bill-curved is not available.

	bill-curved	bill-brown	bill-orange	bill-red
(a) $N_1=0$, $N_2=0$	19.1%	25.6%	38.6%	35.8%
(b) $N_1=4$, $N_2=0$	20.1%	25.9%	39.4%	36.7%
(c) $N_1=24$, $N_2=0$	21.4%	26.7%	42.7%	38.1%
(d) $N_1=24$, $N_2=9$	-	26.9%	43.0%	38.3%

Table 5. Table 6: Average Precision(AP) on our new human attribute test set. The attributes 1-10 are some of the most useful ones in security surveillance applications: 1  the length of sleeve (short / long) , 2  the length of pants (long / short) , 3  using a cell phone or not , 4  carrying an item or not , 5  dragging a luggage or not , 6  smoking or not , 7  wearing a glove or not , 8  holding a baby in arm or not , 9  wearing a mask or not , 10  holding an umbrella or not . The 2nd row shows the number of the training samples. It can be observed that CSN obtains higher performance gain for attributes with fewer training samples.

Attribute ID	1	2	3	4	5	6	7	8	9	10
# of training imgs	51968	12050	4982	3183	1334	542	320	276	209	147
Baseline	89.5%	47.1%	57.0%	38.0%	49.0%	7.1%	5.9%	6.3%	2.2%	0.5%
CSN	96.3%	65.8%	75.6%	67.3%	84.1%	23.4%	32.2%	38.3%	7.9%	30.3%

Table 6. Table 7: Comparison of attribute classification accuracy on CelebA test set.

Baseline	Multi-task [18]	Fully [36]	CSN w/o share	CSN
91.1%	91.2%	91.3%	91.7%	91.8%

Table 7. Table 8: Average Precision(AP) comparison between zero shot learning and supervised learning on 20 attributes. Zero shot is CSN with part and pattern sharing, supervised is CSN w/o sharing.

attribute	#images	zero-shot	supervised
belly_solid	2455	86.1%	88.1%
upperparts_black	1561	74.5%	78.8%
crown_black	1434	75.7%	82.9%
underparts_black	730	67.3%	74.5%
breast_multi-colored	626	30.8%	37.9%
crown_buff	413	38.1%	39.2%
underparts_brown	360	41.2%	47.8%
belly_striped	319	36.5%	41.5%
wing_yellow	316	52.3%	63.9%
belly_brown	313	36.7%	44.8%
wing_spotted	300	37.0%	51.6%
bill_yellow	215	9.2%	50.3%
throat_red	175	65.6%	70.0%
belly_red	140	68.3%	73.1%
wing_olive	127	33.2%	30.6%
upperparts_orange	81	18.7%	28.1%
wing_iridescent	74	12.0%	13.0%
belly_olive	67	13.9%	23.9%
underparts_green	38	25.2%	16.0%
forehead_purple	22	11.2%	4.2%

Equations

$$A_i(x;V_i,\Theta )=A_i(Q;V_i)=Q(x;\Theta ){V_i}^T.$$

$$A_i^{\prime }=s(A_i).$$

$$Q_i^{\prime }=Q(x;\Theta {)}^T\otimes A_i^{\prime }.$$

$$f(x;\Theta ,V_i,W_j^T)=softmax(W_j^T{\bar{Q}}_i^{\prime }).$$

$$L=\sum _{i=0}^{N-1}L(f(x_i;\Theta ,V_i,W_j^T),y_i).$$

$$W_j^{+}=W_j-\gamma \sum _k\sum _{n=0}^{N_{kj}-1}\frac{\partial L(f(x_n;\Theta ,V_k,W_j^T),y_n)}{\partial W_j}$$

$$V_i^{+}=V_i-\gamma \sum _k\sum _{n=0}^{N_{ik}-1}\frac{\partial L(f(x_n;\Theta ,V_i,W_k^T),y_n)}{\partial Q_i^{\prime }}\times \frac{\partial Q_i^{\prime }}{\partial V_i}$$

$$A_k^{\prime }=\sum _{i=0}^{m-1}{t}_i{A}_i.$$

Full text

1 Introduction

The computer vision community has seen tremendous progress in recognizing global features of objects, such as performing category detection [44, 15, 43, 68] and classification [24] (e.g. detect the bounding box and classify the category of a bird from an image). Meanwhile, recognizing attributes of object parts (e.g. localize the wing of a bird and classify its biologic feature) is still a very challenging problem due to multiple issues. First, attributes (e.g. the color of the wing of a bird ) normally attach to a very limited area of an object, which are usually more difficult to be accurately localized from an image compared to the overall object. Most existing part attribute recognition methods [64, 30, 62] train a part detector with large extra annotations to detect the relevant part. However, such part annotations are very expensive to obtain. Therefore, these methods generally fail when the part annotation is not available. How to recognize the part attribute with only image-level annotation is still under-explored.

Another important problem is that the training data is expensive to obtain and usually insufficient in the existing dataset. For example, in a commonly used bird parts attribute recognition dataset CUB-200-2011 [51], the number of training images for most attributes varies from merely a few dozens to at most a few hundred. Most existing attribute recognition methods (if not all) process each part attribute independently, and ignore the spatial correlation of different part attributes. As a result, their performance is simply limited by the volume of training data of each isolated attribute. How to solve the data insufficiency problem is rarely discussed.

To address these challenges, we propose a novel Concept Sharing Network (CSN) for part attribute recognition. In CSN, the part attribute is defined as the combination of two concepts: part location and appearance pattern, as illustrated in Fig. 1. Our neural network models the two concepts as two modules, in contrast to individually modeling each attribute in different branches. Since the two modules in CSN could be shared among different attributes, the labeling of attributes (e.g. color and shape) belong to certain parts (e.g. wing) of an object can be used to facilitate the training of another attribute of the same parts, and vice versa. In such a manner, we maximize the usage of the precious training data to boost the attribute recognition performance independently and aggregately. Note that CSN only needs image-level attributes labels to train, so it would be more general than existing part attribute recognition methods [64, 30, 62] which depend on the part location annotation.

Furthermore, CSN can be used to discover new attributes, i.e. zero-shot part attribute recognition. Given a training set with certain attributes, the training result of part localization and pattern recognition in CSN could be used to recognize a new attribute that does not belong to the training set. As illustrated in Fig. 2, after the wing location and color (red) pattern are learned, a new attribute wing red could be recognized even though the training data of such attribute does not exist.

In this work, we also contribute a large-scale human attribute dataset, named as SurveilA, which contains 75,000 images with 10 carefully annotated attributes focusing on the fine-grained human activities for video surveillance. The human images are collected in the wild under different scenes, scales, poses and viewpoint variations. The dataset is challenging as shown in the experiments that simply fine-tuning standard networks cannot provide accurate enough estimation, and recognition would require a model to focus on local discriminative parts.

Overall, our work has the following contributions:

1. We aim at addressing the data insufficiency problem in part attribute recognition, which is rarely discussed in previous work.
2. We present an effective part attribute recognition framework which does not depend on the part annotation.
3. Our network is proven to be effective in zero-shot part attribute recognition.
4. We will release a new dataset for part attribute recognition, which consists of 75000 images of human in a real-world scenario with 10 attributes annotation.

2 Related Work

Attribute Recognition Attribute Recognition was first introduced as a computer vision problem in [12]. From then, attribute recognition have been studied extensively with numerous datasets and methods [11, 10, 26, 25, 27, 28, 55, 67]. Part Attribute recognition is a harder problem because it is only attached to a very limited area of an object. State of the art methods [5, 64, 30] usually rely on the part location annotations to train part detectors such as Poselets [5], Deformable Part Models [63] or R-CNN [16] to first localize parts then extract visual features to recognize attributes [22]. But the part annotation is very expensive to obtain. Although, recently some methods [57, 21, 69] are proposed to localize the important regions for recognition, they are not carefully designed for part attribute recognition and do not attempt to address the data insufficiency problem.  [14, 32] used attribute recognition results to facilitate the fine-grained recognition, but both of them will fail when training data is insufficient.

Few-shot / Zero-shot Learning Besides collecting more data, few-shot learning [50] and zero-shot learning [39] attempt to directly address the data insufficiency problem - predicting novel concepts that were either very few or completely unseen from the training set. These problems are classical because almost all in-the-wild data follow a heavy-tail distribution [19] with new classes appearing frequently after the training and no finite set of samples can cover the diversity of the real world. Recently, few-shot learning is modeled as a meta learning problem [42, 13] through explicitly building training loss to enforce adaptation to new categories with few examples. On the other hand, due to the complete lack of training data, zero-shot models attempt to learn to transfer knowledge from other external sources [1, 45, 7, 52, 65, 31, 58]. In contrast to these works, we make use of the visual attention mechanism to disentangle part location with appearance features and share the disentangled representations between attributes, which enable us to conduct zero-shot or few-shot generalization on novel attributes.

Visual Attention and Visual Question Answering Visual attention models [38, 4] have been widely used in object recognition [69, 53], fine-grained recognition [47, 33], image captioning [60] and visual question answering [8]. These models also decompose the location and appearance during representation, but do not focus on addressing the data insufficiency problem. CSN improves the visual attention models on data efficiency by sharing the attention mechanism across multiple attribute labels.

The attention region and feature sharing depend on attribute labels, which is similar to visual question answering (VQA) [3]. In VQA, existing methods [2, 37] also try to localize the relevant region according to the given question. But in VQA, all Q/A pairs use the same classifier (i.e. sharing image and language feature extraction, sharing answer classifier). In contrast, recognition of multiple attributes are usually considered as a multi-task problem with different classifiers to be trained. It is very expensive to collect sufficient data so as to well train each classifier, especially in large-scale recognition. This poses a special challenge of data insufficiency, which is the focus of our method.

Attribute Dataset There are some attribute recognition datasets existing, from general objects and scenes [11, 27, 10, 46, 40, 28, 41, 66] to specific fine-grained classes such as faces [20, 26, 25, 35], birds [51], cars [61], clothes [34] or even butterflies [54]. Due to the importance of human attribute recognition [17, 5, 48, 70, 9, 64, 49, 29, 30] in real world applications and challenges, we propose a large scale challenging dataset for human part attribute recognition.

3 Concept Sharing Network

Part attribute recognition aims to predict whether or not the statement of the attribute of a part is true, e.g. whether or not ’the wing of the bird is black’ or ’the bill of the bird is red’. In this work, we introduce a novel Concept Sharing Network(CSN) for part attribute recognition, as illustrated in Fig. 3. In CSN, each attribute is defined as a combination of two concepts: part localization and pattern recognition.

3.1 Network

In CSN, attributes are recognized based on two modules: the part localization module and the appearance pattern recognition module. In our training process, as illustrated in Fig. 2, training images of different attributes are forwarded through the CNN to obtain the image representation, then attribute samples with the same part are forwarded through the same localization module and attribute samples with the same appearance pattern are forwarded through the same pattern recognition module. The modules are learned by the training samples which are forwarded through them. In the inference process, as illustrated in Fig. 3, given an image and recognizing attribute $P_{i,j}$ corresponding to part $i$ and appearance pattern $j$, the image is forwarded through the corresponding location module and pattern module to obtain the final prediction.

3.1.1 Part Localization module

One novelty of our work is the employment of attention mechanism within CSN neural network to localize the part for attributes. We propose an attention based method inspired by CAM [69]. Note that many other alternatives [53, 56] can also be incorporated with our framework. Given an image $x$, we use the stack of CNN to extract features at different location over the image. The output of the CNNs at a particular layer is $Q(x;\Theta)\in\mathbb{R}^{hw\times d}$ where $h$ and $w$ are the spatial height and width respectively, $d$ is the number of channels and $\Theta$ is the parameters of the CNN adopted. For a specific part, we maintain a learnable vector $V_{i}\in\mathbb{R}^{d\times 1}$ which is called part representation to encode the associated part. We expect the inner-product value between $V_{i}$ and features in the feature map $Q(X;\Theta)$ being high in associated regions, while the value being low at other places. As shown in Fig.  3, the localization module takes image representation $Q$ and location representation $V_{i}$ as input and outputs the inner-product map $A_{i}$:

[TABLE]

We then normalize $A_{i}$ over spatial domain by the soft-max operation to derive the attention map as

[TABLE]

where $s$ is spatial Softmax function. The attention map is broadcasted over channels and the results are then multiplied back to the feature map resulting an attention weighted feature map $Q_{i}^{\prime}$:

[TABLE]

where $\otimes$ is the operation of broadcasted multiplication. Note that, the attention weighted feature maps $Q_{i}^{\prime}$ are different for different $i$, though the input feature map $Q$ is the same, which render output features to focus on different spatial location.

3.1.2 Pattern Recognition Module

Conventional attribute recognition methods usually adopt an average pool followed by a fully-connected layer and a soft-max layer to produce final probability of an attribute. In contrast, our method learns an attention map for each attribute to weight the feature map so that the following appearance pattern could be learned by aggregating all training data of different part locations. Specifically, for each appearance pattern $j$ which is annotated a binary classifier label in part attribute recognition, the probability of predicted label is calculated as,

[TABLE]

where ${\bar{Q}_{i}^{\prime}\in\mathbb{R}^{d\times 1}}$ is the global average pooling result of $Q_{i}^{\prime}$ over spatial domain, and $W_{j}\in\mathbb{R}^{d\times 2}$ are appearance specific weights for the binary classification task $j$.

3.2 Concept Sharing Training

In this part, we describe the concept sharing training process. We denote an attribute by $P_{i,j}$ where the part location index is denoted by $i$ and the pattern indexed is denoted by $j$. The attribute recognition model has parameters $\Theta$, $V_{i}$, and $W_{j}$ to optimize. We denote the training images for attribute $P_{i,j}$ as $X_{ij}={x_{0},x_{1},...,x_{N_{ij}-1}}$ whose labels are $Y_{ij}={y_{0},y_{1},...,y_{N_{ij}-1}}$ where $N_{ij}$ denotes the total number of training samples for the attribute $P_{i,j}$. The $\Theta$, $V_{i}$, $W_{j}$ are trained by an end-to-end fashion using cross-entropy loss at final binary outputs of recognition modules. The overall loss is:

[TABLE]

Where $L$ is the cross-entropy loss and $N$ is the total number of training samples satisfying: $N=\sum_{i,j}N_{ij}$. All attributes in Eq. 5 share the same $\Theta$ which are used to extract the CNN features. In our concept sharing training, the localization module is learned from all training samples sharing the same part, and the pattern recognition module is learned from all training samples which sharing the same pattern. Consequently, the weights of sharing models $V_{i}$ and $W_{j}$ are updated by:

[TABLE]

, and

[TABLE]

It is worthwhile to mention that, if the attributes are treated to be independent as in conventional recognition frameworks, training $V_{i}$ and $W_{j}$ will only involve training samples $\left(X_{ij},Y_{ij}\right)$. Since the number of training samples of a single attribute is usually small in practice, and therefore the performance of conventional recognition frameworks is limited by the insufficiency of training samples for each isolated attribute. In contrast, the number of training samples for a particular module would be enlarged in our conceptual sharing framework i.e. $\sum_{k}N_{kj}$ for the pattern recognition, and $\sum_{k}N_{ik}$ for part localization. The training data expansion improves the learning sufficiency with the limited data.

3.3 New Attribute Recognition

In large-scale applications where the number of attributes is large, it is almost infeasible to obtain reasonable amount of qualified training data for every attribute. In this part, we explain how CSN can be used to recognize a new attribute without requiring any training samples. As we discussed above, the localization and recognition module are shared among different attributes, one specific combination of localization and recognition modules determines the recognition for one specific attribute, such as the attribute breast spotted in Fig. 1. Notice that we do not have any training data for breast spotted. However, we can still train breast localization module and spotted pattern module from other attributes. Therefore, by combining the learned location forehead and pattern spotted, forehead spotted could be understandably recognized without any such training data. Specifically, attribute $P_{i,j}$ recognition model is determined by its parameters $\Theta$, $W_{j}$ and $V_{i}$. Since $W_{j}$ could be learned from attributes $P_{\alpha,j}$($\alpha\neq i$), and $V_{i}$ could be learned from attributes $P_{i,\beta}$($\beta\neq j$), attribute $P_{i,j}$ recognition model could be determined even without any $P_{i,j}$ data.

4 Experiments

We conduct experiments on three attribute recognition datasets including CUB-200-2011 [51], CelebA [35] and our new SurveilA dataset. Since the positive and negative samples are highly imbalanced in CUB-200-2011 and SurveilA, we use average precision as our main evaluation metric.

4.1 Datasets

The CUB-200-2011 dataset [51] contains 11,788 images of 200 bird categories, where 5994 images are selected for training and the rest 5794 images for testing. Each image is annotated with 312 attributes among which 278 are part related attributes. These part related attributes are all binary attributes indicating whether a specific appearance pattern exists in one particular part such as attribute wing-black tells whether the wing is black. During experiments,we observe that the back and tail attributes are nosily labelled. Therefore, we exclude these labels and conduct experiments on the remaining 204 attributes. This is the main dataset for our experimental comparison and ablation study.

CelebA [35] consists of 202,599 face images and 40 binary face attributes. 16,000 images are selected for training and 20,000 images for testing and validation.

Our new SurveilA dataset contains 75,000 images with 10 binary attributes, among which 70,000 are selected for training and the rest 5,000 for testing. The dataset focuses on human part attribute (e.g. whether carrying an item, whether wearing shorts or pants), collected under real surveillance scenarios with large pose and appearance variations. We will release the dataset to facilitate the research of attribute recognition.

4.2 Implementation Details

We first resize the image to be $512\times 512$, then randomly cropped $446\times 446$ for training. We use ResNext50 [59] as the visual representation module for feature extraction. The outputs from the layer ‘conv5’ in ResNext50 are used as the feature representation for part localization i.e. $Q$ in eq. 1.The network is trained for 100 epochs with ADAM[23] where initial learning rate is set to 0.0001 with a learning rate decay of 0.1 after 50 epochs.

4.3 Experiments on CUB-200-2011 Dataset

4.3.1 Study the Number of Training Samples

In this part, we empirically study the the effectiveness of CSN as the number of training samples varies. In order to study this, we select attributes belly-solid, breast-white, bill-grey and bill-black which could share the same part and have relatively larger positive samples. All experiments are run on joint training of 91 attributes(i.e. all bill relevant attributes, all grey attributes, all black attributes, all belly, all breast, all white). We evaluate the performance of the CSN w/o(without) share, CSN w/part share, CSN w/part + pattern share on the 3 attributes selected. The results are shown in Tab. 2 and Tab. 2 We see the benefits of sharing attributes is more obvious when the number of training samples is small. The pattern features extracted at different location still varies, forcing them to be the same pattern when they have relatively large data will harm the performance. But when the training data size decreases, their own data is insufficient to learn the pattern module. Pattern sharing will improve the learning efficiency of the pattern module.

We also try to decrease the training samples to zero. CSN still obtains promising results for belly-solid (84.7% AP), breast-white (79.9% AP), bill-grey (46.3% AP), bill-black (64.2% AP) that are comparable to the supervised training. We will further investigate the effectiveness of CSN for zero-shot recognition in the following section.

4.3.2 Study the Number of Sharing Attributes

In this part, we empirically study the effects of the number of sharing attributes on the overall performance. In Tab. 5, we compare sharing different number of attributes in pattern module and localization module. We select attribute bill-curved,bill-brown,bill-orange and bill-red. We perform experiments on joint training of all bill attributes, all brown attributes, all orange attributes and all red attributes. We compare the four attributes performance under different number of sharing part attributes and sharing pattern attributes.

From (a) - (c) in Tab. 5, as the number of sharing part attributes increases, the overall performance steadily increases. This illustrates that it is more effective when number of sharing part attributes become larger. Note that as the number of sharing pattern attributes increases to be 9, the improvement is observed, this also indicates the sharing pattern is effective.

4.3.3 Comparison with the State-of-the-art

We first compare with state-of-the-art recognition method with only image level annotation CAM [69], Spatial Transform Network(STN) [21] and our baseline (a multi task learning framework with different branches). STN aims at explicitly localizing important regions for recognition. We used the public implementation and replace the recognition by our part attribute recognition. Our CAM implementation follows [69] (i.e. localize and crop) for recognition. Tab. 3 shows the performance of our method against other methods. Since the baseline always shows very low performance with attributes with small number of training images, we select the attributes with more than 1000 positive images for a fair comparison. Such selection leads to 32 attributes (e.g. as white relevant attributes and black relevant attributes) in the experiments. CSN produces AP of 65.2% compared to the competitors: 63.2% from CAM, 63.7% from STN and 63.1% of CSN baseline. This illustrates the advantage of CSN over alternatives given a reasonably large training set. We investigate all 204 part attributes to further understand the overall performance of different methods. Fig. 5 shows the AP difference between of CSN and the baseline on all 204 attributes. Significant improvement over the baseline is observed on average.

We then compare with state-of-the-art attributes recognition methods [64, 62]. We used the public implementation [64, 6, 62] to train a parts detector(i.e. poselets [6] and R-CNN [62] ) with part annotation and then recognize the attributes. Without part annotation, we still obtain comparable performance with PANDA [64] and R-CNN based method  [62]. Since they both depend on the part annotations to train the part detector, they fail when the part annotation is not available. Furthermore, they both can not be used for zero shot recognition while ours can.

We also visualize the localization heat map(i.e. the attention map) obtained by CSN w/o share and CSN in Fig.  4. The localization heat map obtained by CSN is obviously better than that obtained by CSN w/o share. This further verifies the effectiveness of the concept sharing.

4.3.4 Soft Sharing among Attributes

In the above section, attributes with the same part are manually grouped to the same part localization module, and attributes with the same pattern are manually grouped to the same pattern module. Accordingly, attributes with different concepts (localization / pattern) can not share the same module. In this part, we study the soft sharing among parts. We add a learnable soft weight vector $T_{i}=t_{1},t_{2},..,t_{m}$ for attribute $k$, where $m$ is number of parts modules. This weight vector is used to combine the attention map from different part. That is

[TABLE]

We first initialize the vector to be one hot vector(i.e. one value is 1, others are 0 in $T_{i}$). Let the attributes sharing the same part have the same initialization, and let them learnable during the training. This obtains 65.1%(i.e. CSN-soft in Tab. 3) which is comparable with CSN. We then initialize the vectors as 1(i.e. all values in $T_{i}$ are initialized as 1). This indicates that we do not have the prior knowledge which attributes are of the same part. We obtain 63.8%(i.e. CSN-soft-1 in Tab. 3) which is lower than 65.1% of the previous one. This indicates that this prior knowledge provide important information.

4.3.5 Zero-Shot Attribute Recognition

In this part, we study the capability of CSN on zero shot attribute recognition. The experiments are carried out on all 204 attributes where we randomly select 20 attributes as unseen attributes and leave the other 184 attributes for training the network. In Tab. 8, zero-shot algorithm refers to CSN w/ sharing pattern and part with no training data on the 20 attributes, this is the algorithm for zero shot in Tab. 8, and supervised refers to CSN w/o sharing trained on all available data. Since the performance gap between zero-shot and supervised is highly relevant to the size of train data, we also list the number of positive samples in statistics. As shown in Tab. 8, zero-shot obtains promising results on the condition of zero shot learning. Note that attributes with very small data, such as ’forehead purple’, ’wing olive’, ’underparts green’, zero shot even outperforms results trained on their own data. For most of attributes, zero shot algorithm is surprisingly effective as it shows accuracy comparable with trained on their own data.

4.4 Experiments on CelebA

Our method can also be applied on the general attributes(i.e. global and parts attributes), so we also perform experiments on a general attributes dataset CelebA [35]. We follow the protocol in [18]. In table 7, we evaluate the average accuracy which is usually reported for CelebA. Our CSN obtains better performance than our baseline and beats the state of the art methods. This is because all other methods fail to explicit localize the important regions for recognition. In CSN, we group the parts attributes to share the same localization module such as nose related and mouth related. We observe further improvement by sharing the localization module. This indicates that the concept sharing is still effective.

4.5 Experiments on SurveilA Dataset

Tab. 6 shows that in this dataset CSN achieves 51.2% mAP compared to baseline mAP 30.3%. Such a huge performance improvement is achieved since most of these human attributes only depend on a small part of the image (e.g. wearing a mask or not are only related to the face area). Hence, the localization function of CSN establishes higher importance than the CUB-200-2011 dataset. In Fig. 7, we visualize the parts localization obtained by our methods. The experiments on humans further verify that our proposed method is effective and reliable for recognizing parts of different types of objects.

5 Conclusion

In this paper, we proposed a new neural network structure (CSN) for part attribute recognition. By identifying part locations and appearance patterns from the training data that do not explicitly label these two concepts, CSN can increase part attribute recognition accuracy especially if the number of labels is small. In the special case of data limitation where none data is available, CSN is still valid to recognize attributes (i.e. zero-shot part attribute recognition).

Acknowledgements

This work was supported in part by National Science Foundation grant IIS-1619078, IIS-1815561, and the Army Research Office ARO W911NF-16-1-0138.

Appendix A Experimental Attributes

In Figure 5, we mentioned that we performed experiments on 204 attributes. The 1-204 attributes are: 1 curved_(up_or_down) , 2 bill_dagger , 3 bill_hooked , 4 bill_needle , 5 bill_hooked_seabird , 6 bill_spatulate , 7 bill_all-purpose , 8 bill_cone , 9 bill_specialized , 10 wing_blue , 11 wing_brown , 12 wing_iridescent , 13 wing_purple , 14 wing_rufous , 15 wing_grey , 16 wing_yellow , 17 wing_olive , 18 wing_green , 19 wing_pink , 20 wing_orange , 21 wing_black , 22 wing_white , 23 wing_red , 24 wing_buff , 25 upperparts_blue , 26 upperparts_brown , * 27 upperparts_iridescent* , 28 upperparts_purple , 29 upperparts_rufous , 30 upperparts_grey , 31 upperparts_yellow , 32 upperparts_olive , 33 upperparts_green , 34 upperparts_pink , 35 upperparts_orange , 36 upperparts_black , 37 upperparts_white , 38 upperparts_red , 39 upperparts_buff , 40 underparts_blue , 41 underparts_brown , 42 underparts_iridescent , 43 underparts_purple , 44 underparts_rufous , 45 underparts_grey , 46 underparts_yellow , 47 underparts_olive , 48 underparts_green , 49 underparts_pink , 50 underparts_orange , 51 underparts_black , 52 underparts_white , 53 underparts_red , 54 underparts_buff , 55 breast_solid , 56 breast_spotted , 57 breast_striped , 58 breast_multi-colored , 59 head_spotted , 60 head_malar , 61 head_crested , 62 head_masked , 63 head_unique_pattern , 64 head_eyebrow , 65 head_eyering , 66 head_plain , 67 head_eyeline , 68 head_striped , 69 head_capped , 70 breast_blue , 71 breast_brown , 72 breast_iridescent , 73 breast_purple , 74 breast_rufous , 75 breast_grey , 76 breast_yellow , 77 breast_olive , 78 breast_green , 79 breast_pink , 80 breast_orange , 81 breast_black , 82 breast_white , 83 breast_red , 84 breast_buff , 85 throat_blue , 86 throat_brown , 87 throat_iridescent , 88 throat_purple , 89 throat_rufous , 90 throat_grey , 91 throat_yellow , 92 throat_olive , 93 throat_green , 94 throat_pink , 95 throat_orange , 96 throat_black , 97 throat_white , 98 throat_red , 99 throat_buff , 100 eye_blue , 101 eye_brown , 102 eye_purple , 103 eye_rufous , 104 eye_grey , 105 eye_yellow , 106 eye_olive , 107 eye_green , 108 eye_pink , 109 eye_orange , 110 eye_black , 111 eye_white , 112 eye_red , 113 eye_buff , * 114 bill_length::about_the_same_as_head* , 115 bill_length::longer_than_head , 116 bill_length::shorter_than_head , 117 forehead_color::blue , 118 forehead_color::brown , 119 forehead_color::iridescent , 120 forehead_purple , 121 forehead_rufous , 122 forehead_grey , 123 forehead_yellow , 124 forehead_olive , 125 forehead_green , 126 forehead_pink , 127 forehead_orange , 128 forehead_black , 129 forehead_white , 130 forehead_red , 131 forehead_buff , 132 belly_blue , 133 belly_brown , 134 belly_iridescent , 135 belly_purple , 136 belly_rufous , 137 belly_grey , 138 belly_yellow , 139 belly_olive , 140 belly_green , 141 belly_pink , 142 belly_orange , 143 belly_black , 144 belly_white , 145 belly_red , 146 belly_buff , 147 wing_rounded-wings , 148 wing_pointed-wings , 149 wing_broad-wings , 150 wing_tapered-wings , 151 wing_shape::long-wings , 152 belly_solid , 153 belly_spotted , 154 belly_striped , 155 belly_multi-colored , 156 leg_blue , 157 leg_brown , 158 leg_iridescent , 159 leg_purple , 160 leg_rufous , 161 leg_grey , 162 leg_yellow , 163 leg_olive , 164 leg_green , 165 leg_pink , 166 leg_orange , 167 leg_black , 168 leg_white , 169 leg_red , 170 leg_buff , 171 bill_blue , 172 bill_brown , 173 bill_iridescent , 174 bill_purple , 175 bill_rufous , 176 bill_grey , 177 bill_yellow , 178 bill_olive , 179 bill_green , 180 bill_pink , 181 bill_orange , 182 bill_black , 183 bill_white , 184 bill_red , 185 bill_buff , 186 crown_blue , 187 crown_brown , 188 crown_iridescent , 189 crown_purple , 190 crown_rufous , 191 crown_grey , 192 crown_yellow , 193 crown_olive , 194 crown_green , 195 crown_pink , 196 crown_orange , 197 crown_black , 198 crown_white , 199 crown_red , 200 crown_buff , 201 wing_solid , 202 wing_spotted , 203 wing_striped , 204 wing_multi-colored.

Appendix B Exemplar images

Following are some exemplar images of our human dataset and CUB bird dataset. The heat map placed on the right of each image visualizes the localization information predicted in the CSN inference process. The heat map is $A_{i}^{\prime}$ in line 299 and Figure 3.