Generating Key Postures of Bharatanatyam Adavus with Pose Estimation

1. Problem Statement

The paper addresses the challenge of digitally preserving Bharatanatyam, a classical Indian dance form characterized by strict structural rules, geometric precision, and symbolic gestures.

• The Core Issue: While general human image synthesis is well-studied, generating Adavu (the fundamental movement unit of Bharatanatyam) key postures is difficult. These postures require exact anatomical alignment (angashuddhi), specific limb coordination, and adherence to cultural norms.
• Limitations of Existing Methods: Standard generative models (like GANs or Diffusion models) prioritize diversity and visual plausibility, often introducing distortions in joint alignment, body symmetry, and limb proportions. Such errors are unacceptable in classical dance, where even minor misalignments alter the symbolic meaning and rhythmic integrity of the pose.
• Data Scarcity: There is a lack of high-quality, annotated datasets specifically for classical dance postures, as existing datasets (e.g., AIST++, NTU RGB+D) focus on contemporary actions and lack the granularity for traditional forms.

2. Methodology

The authors propose a Pose-Aware Generative Framework that integrates a pose estimation module into standard conditional generative models (cGANs and Conditional Diffusion Models).

A. Framework Overview

The system treats the task as a conditional image generation problem: $G: (z, y) \rightarrow I$, where $z$ is latent noise, $y$ is the key posture class label, and $I$ is the generated image.

• Pose Estimation Module: A pre-trained model (MediaPipe) extracts 2D skeletal keypoints from both the generated images and the ground-truth images during training.
• Dual-Loss Supervision: To enforce anatomical and cultural correctness, two specific loss functions are added to the standard generation loss:

1. Keypoint Alignment Loss ($L_{kp}$):

   • Goal: Enforce pixel-level accuracy of joint placement.
   • Mechanism: Calculates the normalized Euclidean distance between corresponding keypoints in the generated image and the ground truth.
   • Formula: $L_{kp} = \frac{1}{K} \sum ||\text{Norm}(P_{gen}) - \text{Norm}(P_{gt})||^2$

2. Pose Consistency Loss ($L_{pose}$):

   • Goal: Preserve internal geometric relationships (symmetry and structure) rather than absolute coordinates.
   • Mechanism: Minimizes the difference in relative distances and angles between joints (e.g., the angle formed by shoulder-elbow-wrist). This ensures the pose remains invariant to global translation or rotation.
   • Formula: $L_{pose} = \sum (R_{gen} - R_{gt})^2$

B. Total Objective Function

The generator is optimized using a composite loss:
$$L_{total} = L_{gen} + \lambda_{kp}L_{kp} + \lambda_{pose}L_{pose}$$
Where $L_{gen}$ is the model-specific loss (Adversarial loss for GANs, Reconstruction loss for Diffusion), and $\lambda$ terms control the weight of pose supervision.

C. Models Evaluated

The authors tested four configurations to determine the best approach:

1. Standard Conditional GAN (cGAN).
2. cGAN with Pose Estimation (CGPE).
3. Conditional Diffusion Model.
4. Conditional Diffusion with Pose Estimation (CDMPE).

3. Key Contributions

1. First Structured Attempt: This is the first work to generate anatomically and culturally faithful Bharatanatyam key postures using deep generative models, addressing the unique spatial and symbolic constraints of Indian Classical Dance.
2. Pose-Aware Guidance Mechanism: Introduction of a novel pipeline integrating pose estimation to enforce keypoint-level and pose-level consistency, ensuring skeletal alignment and joint symmetry.
3. Comprehensive Evaluation: A detailed comparison of GANs vs. Diffusion models with and without pose supervision, demonstrating that pose guidance is critical for culturally constrained generation.
4. Dataset Curation: Utilization of a curated dataset containing 1,276 videos of 10 dancers performing 15 foundational Adavus with 58 distinct variants.

4. Experimental Results

The models were evaluated using Fréchet Inception Distance (FID) (lower is better) and Multi-Scale Structural Similarity Index (MS-SSIM) (higher is better).

• Quantitative Performance:

  • The Conditional Diffusion with Pose Estimation (CDMPE) achieved the best results: FID: 19.32 and MS-SSIM: 0.78.
  • Adding pose supervision significantly improved both cGAN and Diffusion models. For example, cGAN FID dropped from 26.27 to 24.28 with pose estimation, and Diffusion FID dropped from 21.88 to 19.32.
  • The proposed method outperformed existing baselines like InsetGAN, PCGAN, and StyleGAN-Human.

• Qualitative Observations:

  • Models without pose supervision often produced anatomically incorrect limbs or violated the symmetry required by Bharatanatyam.
  • The CDMPE model successfully preserved structural constraints (e.g., Araimandi posture) and symbolic hand gestures (mudras), which are critical for cultural authenticity.
  • The ablation study confirmed that both $L_{kp}$ and $L_{pose}$ are necessary; using only one resulted in suboptimal structural fidelity.

5. Significance and Impact

• Cultural Preservation: Provides a scalable method for the digital documentation and preservation of Intangible Cultural Heritage (ICH), ensuring that the "grammar" of the dance is not lost in translation to digital formats.
• Education and Pedagogy: The generated high-fidelity postures can serve as visual aids for learners, especially those lacking access to expert gurus, by providing accurate references for complex Adavu sequences.
• Generalizability: While focused on Bharatanatyam, the framework is applicable to other codified dance forms (e.g., Kathak, Ballet) and ritual movements that rely on strict geometric alignment and symmetry.
• Future Directions: The authors note that while body posture is solved, generating nuanced facial expressions (abhinaya) remains a challenge, suggesting a path for future research in emotion-aware dance synthesis.

In conclusion, the paper demonstrates that integrating pose estimation into generative models is not just an enhancement but a necessity for synthesizing culturally significant, rule-based human movements, bridging the gap between AI capabilities and traditional artistic integrity.