Learning Modal-Invariant and Temporal-Memory for Video-based Visible-Infrared Person Re-Identification

Visible-infrared person Re-ID handles person retrieval between different modalities, which is implemented by setting RGB cameras and IR cameras. Considering the poor illumination condition in some cases, especially at night, this cross-modal task does enjoy practical significance. Thanks to the image-based RGB-IR person Re-ID dataset constructed by Wu et al. [40], many cross-modal Re-ID technologies have been studied, most of which are based on metric learning[47, 52, 14, 9, 46, 48, 19, 27], feature learning[47, 52, 45, 46, 48, 50, 32], and adversarial learning[6, 37, 39, 4, 29, 4, 29, 34], etc.

As for metric learning and feature learning, Ye et al. [47] extracted multi-modal shareable features in the feature learning stage, after which heterogenous features are projected into a common space and measured by the metric learning. Hao et al. [14] maps the extracted features onto a hypersphere manifold, in which differences between two samples are calculated based on angles. Different from angle measurement in [14], Feng et al.[9] utilized the Euclidean constrain to shrink the cross-modal gap. In [52], inter-modality and intra-modality variations are firstly taken into account, and the discriminative features are then learned through a top-ranking loss. Ye et al. [48] then extended [52] by introducing a bi-directional center-constrained top-ranking loss, further improving the image based person Re-ID performance.

Additionally, the generative adversarial network (GAN) [11] has also been widely applied to cross-modal Re-ID tasks. Dai et al. [6] embedded a discriminator into the network, enforcing the features from two modalities to be unclassified in an adversarial way. In [37], by introducing the CycleGAN [56], an RGB image is transformed to an IR version, while its id-information is preserved. So is the IR image. Furthermore, some researchers [4, 29] also achieved feature learning in an adversarial and disentanglement learning way. Particularly, Choi et al. [4] disentangled id-discriminative features and id-excluded features from cross-modal images, which are then combined to generate modal-different but id-consistent images. However, this strategy encounters a large computational complexity and some generated images with poor quality do give an inferior influence on the performance.

Apart from aforementioned methods, Ye et al. [49] additionally focused on the image properties of RGB/ IR images. For instance, a novel joint learning strategy by channel augmentation and simulating random occlusions is proposed. Moreover, image alignment [32] and pattern alignment [41] are also exploited to alleviate the discrepancies.

Different from image-based person Re-ID, video-based person Re-ID represents a person by a sequence of images, providing temporal-information and a richer appearance[51]. Generally, existing methods mainly adopt RNN[30, 43, 28, 53], temporal pooling (average or weighted)[43, 5, 42, 10], optical flow[30, 53, 2], and 3D convolution[12, 28, 22], etc. For instance, Xu et al. [43] took original person images and corresponding optical flow as the network inputs, so that the motion consistency for a person in different periods is ensured. Then, features extracted from the CNN-RNN module are utilized to compute attention vectors, selecting informative frames over the sequence. A novel Spatial and Temporal Memory Networks (STMN) [8] is proposed, in which features for spatial distractors that frequently emerge across video frames, as well as attentions optimized for typical temporal patterns, are both stored. Furthermore, Aich et al.[1] designed a flexible feature processing module which can be used in any 3D convolutional block for the Re-ID task. Thanks to this module, complementary person-specific appearance and motion information are well captured. Besides, 3D graph convolution is also introduced for video-based Re-ID. Liu et al. [28] employed context-reinforced topology to build a graph, which successfully encodes contextual information and physical information of the human body. By applying the 3D graph convolutional layers to it, spatial-temporal dependencies and structural information are efficiently captured.

Despite the fact that a number of works have been done for person ID, they are only either image-based cross-modal Re-ID or video-based RGB Re-ID. Our HITSZ-VCM is the first dataset combining cross modalities and video data, allowing the study on video-based cross-modal person Re-ID. It not only achieves the 24-hour surveillance, but also gets comprehensive information and obtains much higher Re-ID accuracy.

In this paper, we build the HITSZ-VCM (HITSZ Video Cross-Modal) Re-ID dataset for the video-based cross-modal person Re-ID task. To our best knowledge, this is the first cross-modal Re-ID dataset based on videos. The HITSZ-VCM dataset contains a large amount of images/frames captured by 12 HD cameras with 3840 $\times$ 2160 resolution. Thanks to the modern monitoring technology, all the cameras can automatically shoot both RGB and IR images according to the lighting conditions. Thus, each person is captured by both RGB and IR cameras. Note that all tracklets are processed by an automatic object tracking system, and we then finetune inaccurate annotations manually.

In detail, our HITSZ-VCM dataset contains 927 valid identities. The cameras shoot 25 frames per second and we extract the first frame out of every 5 frames to build the final dataset. According to this setting, every 24 consecutive images are regarded as a tracklet for a person during the same period, and the last frames whose number may be less than 24 form the last tracklet. Totally, there are 251,452 RGB images and 211,807 IR images, which can be divided into 11,785 and 10,078 tracklets, respectively. Of course, the number of frames in a tracklet can also be dynamically set, which is more flexible than many existing video-based datasets. More specifically, 12 cameras are used for our video collection. Generally, most of identities are captured by 3 RGB cameras and 3 IR cameras, and these cameras are non-overlapped.

Our HITSZ-VCM dataset also covers a series of diverse scenarios. Firstly, 7 outdoor, 3 indoor and 2 passages scenes are included. In detail, some common venues like the office, cafe, passageway, playground, and garden are all considered. Besides, each person is captured from multiple angles under each camera, constructing a richer appearance set. Furthermore, some challenging scenarios such as lighting changes (for RGB images), belonging changes, occlusion and viewpoint changes are collected, as displayed in Fig. 2.

Tab. 1 tabulates the comparison between HITSZ-VCM and existing related Re-ID datasets. As we can see, although MARs [54], Duke-Video [42], and LS-VID [21] also enjoy a large number of valid identities, they fail to cover IR images or videos, being incapable for the 24-hour surveillance. In contrast to RegDB [31] and SYSU-MM01 [40], our constructed HITSZ-VCM dataset extends the image-based version to the video-based one, which provides more abundant and valuable information for person Re-ID. Furthermore, there are much more identities captured from more diverse scenarios in our dataset, greatly contributing to the training of the deep network.

In conclusion, HITSZ-VCM enjoys the following characteristics: (1) Constructing the first video-based cross-modal dataset for person Re-ID. (2) Collecting much more valid identities under diverse scenes. (3) Covering challenging but practical cases.

Here we conduct cross-camera and cross-modal retrieval like existing works [40, 54, 21, 31]. In other words, query and gallery are captured by different cameras and modalities. Meanwhile, with respect to the ‘probe to gallery’ pattern, video-to-video matching is adopted to keep consistent with training data. Regularly, we utilize two retrieval modes for HITSZ-VCM: ‘infrared to visible’ and ‘visible to infrared’, to achieve a more comprehensive evaluation. Additionally, we take all the tracklets in one modality as the query set and those from the other modality as the gallery set. Totally, in the ‘infared to visible’ retrieval, there respectively exist 5,159 and 5,643 tracklets in the query set and the gallery set. Vice versa in the ‘visible-to-infrared’ mode. Note that in our implementation, we discard some too short tracklets (less than 12 images).

To quantitatively evaluate the performance on our proposed dataset, the Cumulative Matching Characteristic curve (CMC) and mean Average Precision(mAP) are adopted as the evaluation metrics. Being similar to many methods, we compute the distance scores of all the query features and gallery features to do the ranking work. For testing, cosine similarity is used as the distance measurement.

Here we respectively denote an RGB sequence and an IR sequence as $\mathbf{V}=\left\{\mathbf{V}^{t}\left|\mathbf{V}^{t}\in\mathbb{R}^{H\times W}\right.\right\}_{t=1}^{T}$ and $\mathbf{I}=\left\{\mathbf{I}^{t}\left|\mathbf{I}^{t}\in\mathbb{R}^{H\times W}\right.\right\}_{t=1}^{T}$, where $H$ and $W$ denote the height and weight of the images, $t$ means the $t$-th frame of this sequence, and $T$ is the total number of images in a tracklet. Correspondingly, the ID labels are denoted as $p_{v}$ and $p_{i}$, while $m_{v}$ and $m_{i}$ denote the modal labels.

To transform frame-level features into a sequence-level feature and effectively capture temporal contexts among multiple frames, inspired by [8], we propose a Temporal Memory Refinement (TMR) module. The structure of TMR is shown in Fig. 4. LSTM [16] layers and the SE attention [17] jointly facilitate the exploitation of temporal information, refining the features to enjoy more discriminative information.

Take RGB data $\mathbf{V}$ as an example and the five convolution blocks in our backbone are denoted as ${E}_{res}$. We denote the two sets of frame-level features from ${E}_{res}$ as $\mathbf{f}_{v1}=\{\mathbf{f}_{v1}^{t}\}_{t=1}^{T}$ and $\mathbf{f}_{v2}=\{\mathbf{f}_{v2}^{t}\}_{t=1}^{T}$, which are utilized for the attention weights generation and the frame-level features aggregation, respectively. Features $\mathbf{f}_{v1}$ are first forwarded into two LSTM layers $LSTM^{2}$, so that the temporal context of this tracklet is obtained. By applying a full-connected layer $FC^{t}$ to its associated output from $LSTM^{2}$ and adding $\mathbf{f}_{v1}^{t}$, an attention $\mathbf{a}^{t}$ is obtained by following the SE attention module.

$$\small\mathbf{a}^{t}=SE^{t}\left(\left(FC^{t}\left(LSTM^{2}(\mathbf{f}_{v1})\right)+\mathbf{f}_{v1}^{t}\right)/2\right),$$

(2)

where $SE^{t}$ means the $t$-th SE attention module. Note that $\mathbf{a}^{t}$ denotes the temporal-memory which gives the importance of the $t$-th frame in a tracklet. In other words, $\mathbf{a}^{t}$ plays as the attention weight of the $t$-th frame-level feature.

Based on the aforementioned analysis, the person representation of the $t$-th frame could be refined and the average pooling processing is then utilized to aggregate the frame-level features into a sequence-level one:

$$\footnotesize\mathbf{F}_{v}=\sum_{t=1}^{T}{\left(\mathbf{a}^{t}\odot\mathbf{f}_{v2}^{t}+\mathbf{f}_{v2}^{t}\right)}/T$$

(3)

Similar processing is conducted for IR data with the shared weights, through which its sequence-level $\mathbf{F}_{i}$ is obtained.

Overall, the temporal information is captured by the TMR module, so that the person representations are refined within a single modality. In our training phase, this module is optimized simultaneously with $E_{res}$, which is regarded as a supplementary for $E_{res}$.

After exploiting TMR where the intra-modal features are refined, we then introduce a modal-invariant adversarial learning to remove the gap from two modalities. Referring to the adversarial strategy, AlignGAN proposed in [37] transforms the RGB/IR image to the IR/RGB version in the pixel-level by confusing a modal-discriminator. However, this strategy is constrained on the image generation, which not only increases the computational complexity, but also is quite sensitive to the quality of the generated image. By contrast, inspired by [20], here we achieve the adversarial learning only based on the feature-level, more efficiently getting the modal-invariant features.

According to Eq.(3), the sequence-level features for RGB and IR tracklets are $\mathbf{F}_{v}$ and $\mathbf{F}_{i}$, respectively. Theoretically, if $\mathbf{F}_{v}$ and $\mathbf{F}_{i}$ do enjoy the id-related information but without modal-related information, they cannot be classified to $m_{v}$ or $m_{i}$. To achieve this task, a classifier $W_{m}$ is introduced whose output is a $3\times 1$ vector. Particularly, this output denotes the probabilities of a tracklet belonging to the RGB modality, IR modality, or neither of them. The objective function is formulated as:

$$\small\mathcal{L}_{adv1}\left(E\right)=CE\left(W_{m}\left(\mathbf{F}_{v}\right),m_{3}\right)+CE\left(W_{m}\left(\mathbf{F}_{i}\right),m_{3}\right)$$

(4)

where $E$ is the combination of the backbone $E_{res}$ and TMR, $CE(\cdot)$ denotes the cross-entropy loss, and $m_{3}$ means the third category which is neither belonging to the RGB modality $m_{v}$ nor the IR modality $m_{i}$. To encourage $\mathbf{F}_{v}$ and $\mathbf{F}_{i}$ to enjoy the discriminative information on identities, we also apply the id-related cross entropy loss and triplet loss to them via another classifier $W_{id}$. Thus, the id-related but modal-invariant function can be represented as:

$$\small\mathcal{L}_{id}\left(E,W_{id}\right)=\mathcal{L}_{adv1}+\mathcal{L}_{id}^{ce}+\mathcal{L}_{id}^{tri}$$

(5)

where $\mathcal{L}_{id}^{ce}$ and $\mathcal{L}_{id}^{tri}$ are the id-related cross-entropy loss and triplet loss, respectively.

Of course, the classification capability of $W_{m}$ plays a key role for the modality-invariant feature learning. Here, being similar to existing GAN based methods, the adversarial learning strategy is adopted to additionally update $W_{m}$, shown as follows:

$$\small\mathcal{L}_{adv2}\left(W_{m}\right)=CE\left(W_{m}\left(\mathbf{F}_{v}\right),m_{v}\right)+CE\left(W_{m}\left(\mathbf{F}_{i}\right),m_{i}\right)$$

(6)

In the optimization, Eq.(5) and Eq.(6) are optimized in an alternative way, so that $W_{m}$ enjoying the more strong capacity adversarially contributes to the modal-invariant feature learning. Note that, it is true that some work [35, 13, 6] also learn the id-related features via the adversarial learning in the feature level. However, they are different from each other. As displayed in Fig.5(a), UCDA[35] and MCLNet[13] enforce the probability belonging to each category to be the same. MCLNet deceives the network to confound different modalities while UCDA acts camera-aware domain adaptation. However, a feature which contains both RGB and IR information can also have the same classification result. Obviously, in this case, we are unsure that whether these two modalities are aligned. Referring to cmGAN[6] in Fig.5(b), RGB and IR features are inversely classified to IR and RGB modalities. A limitation is that these two inputs are transformed to only gain each other modal-specific information, while the modal-gap is not measured. By contrast, our used strategy directly classifies different features into an additional class, guaranteeing that they do fall in a same latent space or domain.

Furthermore, for each tracklet with 24 continuous images, $n$ (it can be dynamically selected) images are selected for training. Specifically, 24 images are divided into $n$ parts with $24/n$ images, in which 1 of $24/n$ images is randomly selected to form the training data. Besides, to synchronize RGB and IR tracklets, we shuffle all tracklets first and then select the same number of RGB and IR tracklets batch by batch. When all IR tracklets are selected, we shuffle all tracklets again.

In this subsection, we experimentally analyze the significance of our HITSZ-VCM dataset, as well as the strategies in the proposed method MITML.

As shown in Fig. 6, the values of all metrics increase really slightly when $n$ is relatively large, i.e., 7 and 8. To reduce the time costs in the training phase, we set $n$ to 6 in the following experiments.

Furthermore, as shown in Tab. 3, we evaluate the other two adversarial learning strategies as discussed in Fig. 5. As we can see, the learning strategy in MITML is more effective than that in [35] and [6].

In this section, we further compare our proposed method with existing state-of-the-art visible-infrared cross-modal person Re-ID methods, including DDAG[50], LbA[32], MPANet[41] and VSD[36] and CAJL[49]. Note that, these comparison methods are primarily designed for image-based datasets. For a fair comparison, we conduct an average pooling layer for their generated frame-level features. For those networks whose backbones are ResNet50, including [50, 32, 41, 36], we implement the average pooling after the backbone, which is similar to what we did in our baseline. Furthermore, we pretrain the model for CAJL [49] on AGW [51] before the channel augmented joint learning.

The CMC and mAP obtained by all these approaches on our dataset are listed in Tab. 4. Obviously, our method reports a noticeable improvement than these state-of-the-art imaged-based cross-modal approaches. Specifically, for the infrared to visible retrieval mode, Rank-1 and mAP increase by 7.15% and 3.82% respectively than the second best method CAJL. As for the visible to infrared search mode, Rank-1 and mAP also obtain a significant growth of 4.41% and 4.88%, indicating the effectiveness of our TMR module on the temporal information exploitation.

Based on the above analysis, the importance of video-based cross-modal person Re-ID is proved, and our methods demonstrates a more remarkable improvement compared with existing methods. However, our methods requires a fixed number of images in one tracklet in the training and testing phases, which decreases the flexibility in realistic applications. In our future work, we will aim to design a novel network which can process tracklets with dynamic lengths.