Dance generation from music has become an increasingly attractive research topic in recent years, whereas it is still one of the most challenging tasks. Different from common motions like walking, jumping and sitting that are evaluated by their functionalities, dance is an important form of artistic performance that focuses more on power, fluency and elegance. Most of the previous works formulate the output of the dance generation task as a sequence of motion state parameters [8, 11, 9, 17, 10, 22, 1], e.g., the positions and rotations of joints at each frame. However, in animation industry the motions of characters are mostly animated by key frames and curves [7, 18]. According to the animation principles [20, 7], the curve representation can avoid unnecessary small flickers and make an action fluent. Inspired by these facts, we propose a new perspective to model the dance generation problem as key poses generation and subsequent motion curve generation.
Dance choreography strongly depends on the music, where movements are supposed to be designed based on the rhythm and beat of the song. Furthermore, recent study has demonstrated the co-occurrence of music beats and kinematic beats . Therefore, it is convenient to assume that each key pose of the dance corresponds to a beat of the music and they have an identical occurrence time. Therefore, we apply a beat track algorithm  to detect the beats of a music, and we generate key poses based on the music spectrum features on the windows centered at the beats, where the poses are parametrically formulated as joint translation and rotation
in the form of quaternion. Given a sequence of key poses, the motions between adjacent key poses are parametrically represented as Cubic Hermite Spline that interpolates the key poses with respect to time. The key benefit of using spline is that it enables various movements based on a small set of control parameters, which can be learned efficiently through a deep neural network.
To this end, we propose DanceNet3D, a two-stage framework with encoder-decoder architectures and adversarial training scheme. The encoders take the well-known transformer  as the basic models. While for the decoders, we propose a novel architecture named MoTrans, that is also based on transformer but has carefully been tailored to the dance generation task. Specifically, in MoTrans we propose the Kinematic Chain Networks (KCN) as the backbone, and the Learned Local Attention (LLA) module with structured multi-head attentions to replace the vanilla attentions.
Inspired from the kinematics that is widely used in robotics and motion control, we propose the Kinematic Chain Networks (KCN) with joints as nodes and bones as edges to model the spatial correlation of body parts. It encodes the features into the kinematic space and conducts feature fusion for transformation controlling forward-and-backward along the kinematic chains. The KCN involves spatial correlation and physics constraints into the model, which leads to more natural results.
Different from sentences in NLP that have long-term correlations, human poses and motions only have local correlations in a short time interval. That is, current pose is strongly related to the neighboring ones, but it only has weak correlation to the poses before or after many seconds. Therefore the global attention in the original transformer is not so reasonable in this task. And we introduce kernels in the attention function and propose the Learned Local Attention (LLA) module with the use of learned kernels. Owing to the LLA module, the training is more stable and the degradation to averaging actions is avoided.
Moreover, we propose PhantomDance, the first 3D dance dataset produced by professional animators. Existing 3D dance datasets are either by reconstruction from 2D videos [8, 11, 9] or by motion capture [22, 17]. The reconstructed data can not ensure the accuracy due to the limitation of recent 3D reconstruction algorithms. The motion capture methods are more accurate but still suffer from noise and misalignment to music. In contrast, our PhantomDance is produced by a team of experienced animators with the help of an expert dancer. The animated dances are more fluent, elegant and expressive, and they are strictly matched with the music rhythm. We will release 100 of the professionally animated dance data with aligned musics as PhantomDance-100 dataset to facilitate future studies.
Extensive experiments are conducted on PhantomDance and AIST++ , which demonstrates that DanceNet3D attains state-of-the-art results and significantly surpasses other works both qualitatively and quantitatively.
Above all, the contributions are summarized as follows:
We propose a new perspective to model the 3D dance generation as key poses and motion curves generation. This formulation takes animation principles into consideration and fits well with the fluency and elegance of dance performance.
PhantomDance is the first dance dataset that crafted by professional animators. Since existing datasets collect dance data either by 3D reconstruction from 2D videos or using motion capture devices, PhantomDance is unprecedented. The dances in it are much more fluent, expressive, and synchronized well with music.
We propose a novel framework named DanceNet3D, and the MoTran architecture as well. MoTrans consists of Kinematic Chain Networks (KCN) and Learned Local Attention (LLA) modules with structured multi-head attention. It introduces kinematic information and temporal locality into the networks to better handle the dance generation task.
2 Related Work
Dance Datasets: Large datasets are the cornerstone of recent advances in motion synthesis using deep learning. The widely used motion datasets, such as Human3.6M  and AMASS , collect large amount of common actions like walking, running, jumping and sitting. However, these datasets are hard to use for dance generation due to the limited motion data distribution and the absence of music-dance pairs. In reality, 3D dance movements of high quality with synchronized music are extremely hard to collect and many of existing datasets are limited by quantity or quality of data. Some of them just use 2D keypoints but no 3D data [8, 9]. AIST++ , as the largest dataset to date, presented a 5 hours long 3D dance dataset. But it also reconstructs 3D poses from 2D multi-view videos, which has limitation of accuracy especially for the rotation parameters. Other works use motion capture to build dataset [1, 17, 22]. The accuracy is ensured, but they suffer from the problem of dis-alignment between dance-music pair. The proposed PhantomDance is the first 3D dance dataset crafted by professional animators, which is high-quality, expressive, graceful and strictly aligned with music.
. Recent years methods based on deep learning attain more and more qualitatively pleasing results. The capability of deep neural networks to generate dances has been demonstrated by many works, such as those using long-short term memory networks (LSTM)
, convolutional neural networks (CNN)[9, 22] and variational auto-encoder (VAE) . CSGN  generates the entire dance sequence altogether with a graph neural network (GNN). The GNN in CSGN grows from root node to leaf nodes, while Kinematic Chain Networks proposed in this work adopt forward and inverse kinematic passing that better involves the physics constraints. And the generated dance of CSGN has no relation to music. As transformer  achieves great success in sequence-to-sequence generation tasks in language processing field, some of the most recent research works on the dance generation also bring it into use. The most related works to ours are those using transformers. Jiaman Li  presents a two-stream motion transformer generative model. And AI Choreographer  presents a cross-model transformer with future-N supervision mechanism for auto-regressive motion prediction. Both of the two works still use fully-connected layers for feature embedding and make no changes on the attention operation. While we propose Kinematic Chain Networks to embed the kinematic information and Learned Local Attention module to enable the model to adapt to the temporal locality of motions. Moreover, all the related works above take the frames of joints parameters as the output of the model, whereas we propose a new formulation that model the output as key poses and the in-between motion curves.
3 Data Collection and Formatting
3.1 PhantomDance Dataset
We first collect 300 popular dance videos on YouTube and NicoNico. Their categories are mainly in Otaku Dance, Hip-hop, Jazz and Old School. Then a team of experienced animators make the dance animations according to the collected music and videos. An expert dancer is also asked to provide instructions on dance acting for the animators, so that the animated dances are graceful and strictly matched with music rhythm. It costs about 18 months for the animators to produce the 300 dance animations, which have 12 hours (about 2.6M frames) in total. These data are named PhantomDance, and we will release 100 of them as the PhantomDance-100 Dataset, which has 794776 frames (about 3.7 hours).
The dances are animated in Maya  on a standard SMPL  character model with 24 body joints. Although the animations in Maya are edited by key-frames and curves, they are exported as 60-FPS pose sequences for convenient use of the dataset. A pose of character is represented as global position and rotation (in quaternion) for each of the 24 joints.
3.2 Key Pose Extraction
We first leverage the a dynamic programming algorithm  to track the time of each beat in the music. Previous study shows that music beats and kinematic inflexion points have strong consistency in time , and moreover our dance data are highly matched with rhythm. So we use the poses corresponding to the beat times directly as the key poses. A key pose is represented as position and rotation (in quaternion) of each joint, with the dimension of .
3.3 Motion Curve Formulation
On the basis of key poses, the dance is split into several pieces. And the in-between motions are pose frames in original dataset. Our target is to model the motions as curves on the transformation parameters (x, y, z for position and x, y, z, w for rotation) of joints. Thus a motion in-between consists of 24 x 7 curves, and each curve has two axes, namely the horizontal axis of time and the vertical axis of parameter value. And we use a multi knots cubic Hermite spline  to fit each curve. It is simple in parametric form and can keep smooth at endpoints, so it is widely used in industrial design including animation.
Cubic Hermite spline (CHS) uses knots (endpoints) and their tangent (slope) to represent a curve. Let be the coordinate of the i-th knot and be its slope. The blend form of the CHS between and is:
where is the uniformed time. and are slopes scaled by the interval length. The coefficient functions are specified in Hermite forms:
In implementation we use a 4-knot CHS to fit a motion curve between two key poses, as an optimal tradeoff between fitting precision and representation simplicity. For a 4-knot CHS, we have 12 control parameters, namely and , , to determine. While and
are just the points corresponding to the two key poses on both ends of the motion curve. So there are 8 control parameters to compute (which are also the output target of the neural networks in stage 2). Since the data in PhantomDance have a 60-FPS format of discrete poses, we just use the neighbor frame points on the two sides of a knot to estimate its slope. Then we setand to the two frame points that minimize the fitting error in the frames between the two key poses knots.
As illustrated in Fig.1, the networks in stage 1 (S1) generate a sequences of key poses based on the input music. And the networks in stage 2 (S2) predict the motion curves between the key poses with regard to the corresponding music. The encoder adopts a standard transformer following . While we propose the MoTrans architecture in the decoders of both stages. And adversarial learning is applied during the training in both stages. Details are specified in the following sections.
4.1.1 Key Pose Generation
As described in Sec.3.2, we obtain the beat times by an off-the-shelf algorithm . Our target in this stage is to generate a sequence of key poses at the same times of the beats. For each beat, we use a hamming window of 0.5s length centered at it and calculate MFCC  on the corresponding wave data. We adopt the 40-dimension form of MFCC that has 13 coefficients of Mel filters cepstrum coefficients, their first and second order differences (13 x 2 dimensions) and the energy sum (1 dimension). We further append a 12-dimension chroma feature to every beat features and finally attain the inputs to the encoder.
As shown in Fig.2, the encoder takes beat features as input and utilizes common self attention modules of the transformer. The attention modules enable the encoder to extract global features of the music (e.g. music style) across beats. The decoder takes right shifted outputs to perform self attention and then combines the features from encoder to perform vanilla attention. In S1 the start parameters refer to a given start pose before all the beats, which can be considered as an opening pose for the dance. During training the start pose is simply selected as the first frame of the dance sequence and the right shifted outputs are just taken from ground-truth with mask for future time steps. While in the test phase any reasonable pose can be used as the start pose, which brings out diversity for dance generation. Note that the network architecture in the decoder is MoTrans that consists of the proposed Learned Local Attention module and Kinematic Chain Networks structure. Details of MoTrans are specified in Sec.4.2.2 and Sec.4.2.1.
The encoder-decoder networks are supposed to map a distribution in the music spectrum space to a distribution in the body pose space. So we adopt an adversarial training architecture that regards the encoder-decoder as a generator. And the discriminator is aimed to determine whether the given key poses are suitable for the input music as well as authentic in human view. Thus we use a 3-layer fully-connected network to embedding the encoded music feature at each time in the sequence, and the embedded features are then average-pooled across the time dimension to obtain the global music features. A similar 3-layer network is employed to extract the global features of the generated/ground-truth key poses. The music and key pose global features are then concatenated and fed into a 3-layer MLP, and the networks finally output a probability of the authenticity of the key pose sequence conditioned on the music.
Let and represent the generated and ground-truth key pose sequence respectively and be the function of discriminator networks. We apply the generally used binary cross entropy loss for the discriminator. Thus the adversarial loss for generator is:
where is the authenticity label (1 for ground-truth and 0 for generation). And a L-2 loss of reconstruction is additionally applied to the generator:
So the final loss function for the key pose generator is:
where is the weight to control the contribution of the loss term.
4.1.2 Curve Prediction
Based on the generated key poses, networks in S2 are targeted to generate the curves in between. S2 has a similar adversarial training schema and network architecture as S1, as shown in Fig.2. For a curve to predict, we extract MFCC features from the music wave data between the two beats corresponding to the two key poses on the ends of the curve. Similar to S1, the 40-dimension MFCC features with 12-dimension chroma information are fed into the encoder. The decoder in S2 outputs 8 control parameters specified in Sec.3.3, and it has a similar right shifted output curve parameters for self attention, except that the output curve parameters are concatenated with the two generated key poses (shown as the extra parameters in Fig.2). The loss function of the curve generator has the same formulation as that of the key pose generator.
4.2 MoTrans Architecture
Both the decoders in S1 and S2 are built with the MoTrans architecture. MoTrans has a similar construction to the original transformer as illustrated on the right of Fig.2. The proposed Kinematic Chain Networks (KCN) is used for output embedding and the fead-forward part in the layer groups. And the two attention sub-layers of layer groups adopt the proposed Learned Local Attention (LLA) modules in the form of structured multi-head attention. We employ a stack of layer groups in MoTrans.
4.2.1 Kinematic Chain Network
The Kinematic Chain Networks (KCN) are constructed on the basis of body structure, i.e., the 24 joint nodes with a tree topology. The networks consist of a stack of core blocks (FK-IK block) as illustrated in Fig.3.
The core block has two steps. It first performs a feature embedding on the root node and pass it to the neighbour child nodes along the kinematic chain. Receiving the features from the parent node, the child node performs feature fusion and then conducts feature embedding by linear projection (FC). This process is named forward kinematic feature passing, where fully-connected layers (FC) are utilized for feature embedding and the fusion adopts the addition operation. Then a similar feature embedding and passing process is performed from the leaf nodes to the root, completing the inverse kinematic feature passing. These procedures are similar to the operations of forward kinematics (FK) and inverse kinematics (IK), thus we name it as FK-IK block. The FK-IK block attempts to encode features into the motion controlling parameter space and to bring physics constraints into the model.
For the output embedding process, we use a KCN with one FK-IK block. And for fead-forward process in the core layers, we adopt a stack of 2 FK-IK blocks for KCN. All the KCNs have the same output feature size of 64 in implementation.
4.2.2 Learned Local Attention
To enable the networks to adapt to the locality property of motion states (joint parameters), we propose the Learned Local Attention (LLA) module. The generally used attention function proposed in the original transformer  is
where represent the matrix of queries, keys and values with a sequence length of
. Query vectors and key vectors have the same feature dimension, and is the feature dimension of value vectors. We use represents to focus on the effects of Q, K and V. Then the attention function for a certain query vector is
Fig.4 (left) provides an illustration of the mechanism of the kernel function. Here we introduce the kernel function that acts on the value vector sequence:
In this equation, is the index filter operation to select out the keys in the range of , because key and value vectors are pairwise and must have a same sequence length. The generally used attention in Eq.7 can also be regarded to have a uniform weighted kernel whose size is just the sequence length . The kernel of the proposed LLA is centered at time step with learnable parameters to weight the contribution of value vectors. The kernel size is set to 17 so as to involve the neighboring 8 beats before and after current beat in consideration of the dance principles.
The kernel parameters are shared across a sequence (the time dimension
), so there are 17 parameters for one LLA module to learn. In the training phase, the kernel functions are initialized with a Gaussian distributionwhere with the assumption that nearer motion state has stronger influence.
4.2.3 Structured Multi-Head Attention
Since features output by KCN are arranged in the joint nodes dimension, they are naturally compatible with the ”multi-head” mechanism used in common transformers . To maintain a unified network structure, the encoded features of the input music, which is used as V, K for the vanilla attention module (the upper LLA part in Fig.2), are processed by 24 paralleled fully-connected layers with the output size of 64 to obtain joint node features with the same structure.
With the nodes as ”heads”, we propose the structured multi-head attention process as illustrated on the right of Fig.4. For each node, 3 individual fully-connected layers embed the joint node features into Q, K and V vectors and then we perform 24 individual LLA modules on Q, K and V of the nodes. Note that no concatenation is needed because the features should maintain the body structure to be fed to the succeeding KCN. Thus this process actually saves computation compared to the typical multi-head attention that concatenates features from each head and use a big linear layer for feature embedding.
The structured multi-head attention briges the KCN and LLA modules, and the head here is the node in the body structure which has explicit physical significance. To some degree this design is more reasonable than the multi-heads in the original transformer.
The experiments are performed on our PhantomDance dataset and the AIST++  dataset.
|Li et al. ||16.31||5595.91||3.40||0.019||0.359||18.34||7944.78||5.84||0.014||0.175|
|AI Choreographer ||8.29||113.56||0.45||0.509||0.452||10.62||164.33||0.73||0.624||0.388|
5.1 Implementation Details
Since most songs of the collected dances have the verse-chorus form, we divide the 300 dance animations into 1092 pieces according to their music structures. Among them 1000 pieces of music-dance pairs are used as input-supervision sequence pairs for model training, and the other 92 are split into test set.
The DanceNet3D is end-to-end trained using 4 TITAN Xp GPUs with a batch size of 8 on each GPU. We use Adam optimizer with betas (0.5, 0.999) and a learning rate of 0.0002. The learning rate drops to 2e-5, 2e-6 after 100k, 200k steps. The model is trained with 300k steps for AIST++ and 400k steps for PhantomDance.
5.2 Evaluation Metrics
Normalized Power Spectrum Simularity (NPSS)
is a better evaluation metric for long-term motion synthesis compared to Mean Square Error (MSE). We just follow the official implementation in and compute NPSS in the joint motion space (4 for joint rotation represented as quaternion and 3 for joint position).
Frechet Distance (FD) is proposed by AIST++  that has two metrics, namely PFD for position and VFD for velocity. We also employ it to calculate the distribution distance of joints.
Position Variance (PVar)
Position Variance (PVar)evaluates the diversity of the generated dance. In AIST++, a piece of music corresponds to more than one dance, but only one in PhantomDance. So we make a modification to the metric PVar in the experiments on PhantomDance where we compute it along different music pieces which have identical length.
Beat Consistency Score (BC) is a metric for motion-music correlation. We follow  to define kinematic beats as the local minima of the kinetic velocity. BC computes the average distance between every music beat and its nearest kinematic beat with the following equation:
where are kinematic beats and music beats respectively and . Note that BC has a very similar form to Beat Alignment Score (BA) proposed in , but they are different in essence. BA forces every kinematic beat to match a music beat, but in fact a dance usually has many small kinematic beats that occur between beats. Moreover a music synchronized dance just need to ensure the most music beats are accompanied with action emphasis (kinematic beats). So our proposed BC, which finds matched kinematic beat for each music beat, is a better metric.
|Baseline w/o Curve||15.11||678.49||2.97||0.343||0.327|
5.3 Ablation Study
To evaluate the impact of the core ideas in DanceNet3D, we conduct the ablation study on our PhantomDance dataset. We use the typical transformer model  with 24 heads attention (same number as ours for fair comparison) as the baseline to compare with our MoTrans. The results are shown in Table 2.
Effectiveness of Curve Prediction: To evaluate the effectiveness of curve prediction, we perform experiments on the method that uses frame by frame generation for comparison, which is a common practice in previous works. The result is shown in the table as Baseline w/o Curve. The experimental result demonstrates the advantages of curve representation. Especially the improvement on the beat consistency score reveals the fluency property brings by motion curves.
Effectiveness of KCN: To study the impact of the KCN component, we introduce a comparison architecture which substitutes the FK-IK block with a 2-layer trivial graph neural network(GNN) . This GNN fuses the joint features according to the adjacency matrix of undirected graph defined by the joint structure. The experimental result shows that GNN has better perforamce compared to the baseline, which proves that the representative ability of the model can be improved by introducing spatial information. While the proposed KCN surpasses the trivial GNN by a large margin. We claim that the proposed KCN, which involves spatial correlation and physics constraints into the model, brings significant improvements especially in generation quality (NPSS, PFD and VFD).
Effectiveness of LLA: Then we evaluate the effects of the LLA module. To verify the contribution of the learned kernel, we add experiments which applies a fixed Gaussian function with kernel size of 17 and standard variance of 5 for local attention. The results are shown as +Gauss, which proves the introducing of temporally local attention mechanism has improvements on the performance for the dance generation task. While the utilization of LLA module (the line of +LLA) demonstrates that a learned kernel can attain rather more gains. This is mainly owing to the rich representitive ability brought by the LLA module, which enables the kernel to adapt to the specific feature distribution. We argue that concentrating on the adaptive local motion information makes the model generate more real transitions.
Finally, we perform the experiments on the complete model, which produces the optimal results in our study. It demonstrates that the kinematic controlling information introduced by KCN and the temporal locality property brought by LLA contribute in different aspects and altogether can improve the model greatly.
5.4 Comparison with Other Methods
We mainly compare our method with AI Choreographer  which to our knowledge obtains current state-of-the-art results for dance generation. Two other most related works, namely Li et al.  and Music2Dance  are also compared in this section. The comparison experiments are performed on both the AIST++ and our PhantomDance datasets.
The results are viewed in Table 1. Our method outperforms Li et al.  and Music2Dance  by a considerably large margin. And it also surpasses AI Choreographer  significantly in the quality related metrics (a 30% gain of NPSS, a 28% gain of PFD and a 27% gain of VFD), diversity metrics (a 46% gain of PVar) and beat consistency metrics (a 93% gain of BC).
The excellent results on the quality results are mainly owing to the appropriate architecture of KCN and LLA. While the generation diversity is ensured by the adversarial learning schema. Note that the upper bound of the metric BC is 1, which means there exactly exists a kinematic beat at the time of each music beat. So our results have relatively high beat consistency score, which is due to the two-stage framework of DanceNet3D.
Since no metric emphasizes the fluency of the dance performance, the advantages of motion curve formulation can only be well revealed in qualitative results.
5.5 Qualitative Results
Fig.5 provides a visualization of generated dances. Results in video form including comparison with other works can be found in the supplementary materials. Owing to curve representation, our results more fluent and graceful.
We also perform a user study to evaluate the authenticity and beat consistency of generated dances conditioned on music qualitatively. We first select 50 pieces of music from the validation set of PhantomDance. And then we collect the generated dances based on the music by conducting our method and three other works [11, 10, 22]. In addition, the ground-truth dances are also included. The user study is conducted using a pairwise comparison schema. For each of the 50 pieces musics, we provide 4 pairs in which our result occurs with result of other method or the ground truth. The pairs are provided to the participants, and they are asked to make two choices for each pair: “Which dance is a better performance (more fluent, graceful and pleasing) regardless of music?” and “Which dance is more synchronized when played with the music?”. There are 100 participants in the user study.
Fig.6 shows the user study results, where DanceNet3D outperforms the other methods on both criteria. Most of the participants think that our method generates better dances in performance quality compared with other works, owing to the curve representation to ensure the performance fluency. And even more participants hold the opinion that dances generated of our model are more synchronized with music. Compared to the ground truths (real dances), there are still 29% of users prefers our approach in quality and 38% in suitability for music. That is, to a great extent, a demonstration of the effectiveness of DanceNet3D.
In this work, we propose a new perspective to model the 3D dance generation task. Different from previous works that define the outputs as frames of poses, we formulate them as key poses and in-between motion curves. The curve representation makes the generated results more fluent and graceful. Based on the formulation we propose a dance generation framework, DanceNet3D with the novel MoTrans architecture consisting of the KCN and LLA module. KCN involves kinematic controlling information and LLA enables the attention function to focus on the nearby time steps that have strong influence on current time. And the structured multi-head attention naturally enriches the feature expressions and bridges the KCN and LLA. Due to the careful design of MoTrans, DanceNet3D yields high-quality results in experiments. Moreover, we propose PhantomDance-100 Dataset, the first 3D dance dataset proposed by professional animators. Future works based on the high-quality PhantomDance-100 can pay more attention to the artistic performance.
-  (2017) Groovenet: real-time music-driven dance movement generation using artificial neural networks. networks 8 (17), pp. 26. Cited by: §1, §2.
-  (2004) Motion retrieval and its application to motion synthesis. In 24th International Conference on Distributed Computing Systems Workshops, 2004. Proceedings., pp. 254–259. Cited by: §2.
-  (2012) Introducing autodesk maya 2013. John Wiley & Sons. Cited by: §3.1.
-  (2007) Beat tracking by dynamic programming. Journal of New Music Research 36 (1), pp. 51–60. Cited by: §1, §3.2, §4.1.1.
-  (2019) A neural temporal model for human motion prediction. In , pp. 12116–12125. Cited by: §5.2.
-  (2013) Human3. 6m: large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence 36 (7), pp. 1325–1339. Cited by: §2.
-  (1987) Principles of traditional animation applied to 3d computer animation. In Proceedings of the 14th annual conference on Computer graphics and interactive techniques, pp. 35–44. Cited by: §1.
-  (2019) Dancing to music. Advances in Neural Information Processing Systems 32. Cited by: §1, §1, §1, §2, §2, §3.2.
-  (2018) Listen to dance: music-driven choreography generation using autoregressive encoder-decoder network. arXiv preprint arXiv:1811.00818. Cited by: §1, §1, §2, §2.
-  (2020) Learning to generate diverse dance motions with transformer. arXiv preprint arXiv:2008.08171. Cited by: §1, §2, §5.4, §5.4, §5.5, Table 1.
-  (2021) Learn to dance with aist++: music conditioned 3d dance generation. arXiv preprint arXiv:2101.08779. Cited by: §1, §1, §1, §2, §2, §5.2, §5.2, §5.4, §5.4, §5.5, Table 1, §5.
-  (2000) Mel frequency cepstral coefficients for music modeling.. In Ismir, Vol. 270, pp. 1–11. Cited by: §4.1.1.
-  (2015) SMPL: a skinned multi-person linear model. ACM transactions on graphics (TOG) 34 (6), pp. 1–16. Cited by: §3.1.
-  (2019) AMASS: archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 5442–5451. Cited by: §2.
-  (2008) The graph neural network model. IEEE transactions on neural networks 20 (1), pp. 61–80. Cited by: §5.3.
-  (2010) Retrieval and generation of human motions based on associative model between motion symbols and motion labels. Journal of the Robotics Society of Japan 28 (6), pp. 723–734. Cited by: §2.
Dance with melody: an lstm-autoencoder approach to music-oriented dance synthesis. In Proceedings of the 26th ACM international conference on Multimedia, pp. 1598–1606. Cited by: §1, §1, §2, §2.
-  (1995) The illusion of life: disney animation. Hyperion New York. Cited by: §1.
-  (2017) Attention is all you need. arXiv preprint arXiv:1706.03762. Cited by: §1, §2, §4.1, §4.2.2, §4.2.3, §5.3, Table 2.
-  (2012) The animator’s survival kit: a manual of methods, principles and formulas for classical, computer, games, stop motion and internet animators. Macmillan. Cited by: §1.
-  (2019) Convolutional sequence generation for skeleton-based action synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 4394–4402. Cited by: §2.
-  (2020) Music2dance: music-driven dance generation using wavenet. arXiv preprint arXiv:2002.03761. Cited by: §1, §1, §2, §2, §5.4, §5.4, §5.5, Table 1.
-  (2020) Advanced engineering mathematics. Jones & Bartlett Publishers. Cited by: §3.3.