MNManish Narwaria

Manish Narwaria

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article

Previous Organizations / Institutes

The organizations / institutes from the scholar's other articles in the past ten years


Manish Narwaria has a total of 48 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    On the Representation Limits of Facial Blendshapes for Micro-Expression Spotting

    by Suraj Sundaram, Manish Narwaria

    MediaPipe’s 52 facial blendshape coefficients have become the de facto representation for real-time facial analysis and avatar animation, yet their suitability for continuous micro-expression (ME) spotting remains unknown. We present the first systematic evaluation of ME spotting in the blendshape domain through a representation evaluation framework comprising a training-free, fully invertible vertex–time spectral graph wavelet transform (VT-SGWT), matched filtering, controlled validation, and an oracle-derived representational upper bound. The proposed framework achieves an ME-only F1-score of 5.13%, while the oracle attains an AUC of only 0.691 despite perfect temporal localization. Together, these analyses consistently indicate that the observed performance ceiling arises from the limited discriminative information preserved by the standardized 52-dimensional blendshape representation rather than from the spotting algorithm itself. The findings establish the representational limits of deployment-oriented blendshape representations for continuous micro-expression spotting and provide a principled framework for evaluating standardized facial representations for downstream affective computing applications.

  • Open Access

    On the Representation Limits of Facial Blendshapes for Micro-Expression Spotting

    by Suraj Sundaram, Manish Narwaria

    MediaPipe’s 52 facial blendshape coefficients have become the de facto representationfor real-time facial analysis and avatar animation, yet their suitability for continuousmicro-expression (ME) spotting remains unknown. We present the first systematicevaluation of ME spotting in the blendshape domain through a representationevaluation framework comprising a training-free, fully invertible vertex–time spectralgraph wavelet transform (VT-SGWT), matched filtering, controlled validation, and anoracle-derived representational upper bound. The proposed framework achieves an ME-onlyF1-score of 5.13%, while the oracle attains an AUC of only 0.691 despite perfecttemporal localization. Together, these analyses consistently indicate that the observedperformance ceiling arises from the limited discriminative information preserved by thestandardized 52-dimensional blendshape representation rather than from the spottingalgorithm itself. The findings establish the representational limits of deployment-orientedblendshape representations for continuous micro-expression spotting andprovide a principled framework for evaluating standardized facial representations fordownstream affective computing applications.

  • Deep learning–based image super‐resolution in microscopy: Why more pixels do not imply higher resolving ability?

    by Manish Narwaria

    Deep learning–based image super‐resolution is a popular topic in computer vision and artificial intelligence (AI)–based imaging. A survey of literature reveals that it is also increasingly being used in the domain of microscopy. From the fundamental perspective of sampling, image super‐resolution is clearly understood and implemented as signal interpolation. However, the widespread use of deep learning–based image super‐resolution in microscopy has led to some misinterpretations regarding the fundamental nature and purpose of image super‐resolution. This is particularly problematic in the domain of microscopy because image super‐resolution may be erroneously considered equivalent to improving resolving ability (i.e. better distinction between finer visual details) of the processed image. We, therefore, rely on first principles of imaging to provide theoretical analysis into why deep learning–based image super‐resolution cannot, in general, reconstruct uncaptured visual details in the context of microscopy. In addition, we provide analysis of the case where deep learning–based image super‐resolution may be useful for the purpose of emphasizing high‐frequency details. The rigorous analysis and conclusions reported in the paper are both relevant and timely as they provide grounded insights towards correct interpretation of deep learning–based image super‐resolution in the context of microscopy.

  • Top Cited Papers

    Adaptive self-Attention enhanced conditional GAN for image dehazing

    by Priyanka Saxena, Anil Kumar Tiwari, Manish Narwaria


  • Haptic-Sound Analysis of Materials’ Clustering Based on Tool Tip Tapping Exploration

    by Kriti Datta, Amit Bhardwaj, Manish Narwaria

    Tapping surfaces with a tool tip elicits both sound and haptic information. Sound and haptic information is captured by a microphone and an accelerometer attached to the tip, respectively. In relation to the task of distinguishing objects, this paper investigates the following two questions: (1) how do both the signals (sound and acceleration) help us to assess the surface individually? (2) How does the integration of both modalities affect this assessment? We approach the problem of texture assessment as an unsupervised learning problem. For this purpose, perceptual filter banks are designed based on Weber’s law of frequency perception for both modalities to extract the corresponding features. Furthermore, we introduce a symmetric KL divergence-based texture similarity metric, which helped us to compare the sound and haptic (acceleration signals) modalities. Based on our similarity metric comparisons, we argue that proximity between object types is preserved across modalities. Finally, using a permutation test-based approach, we demonstrate that the complementarity of sound domain information to the haptic domain varies depending on the type of object.

  • Does Attribute Space-Driven Tactile Features Correspond to Overall Perceptual Similarity of Textures?

    by Kriti Datta, Amit Bhardwaj, Manish Narwaria

    In the recent past, various hand-crafted tactile features have been extracted from object surfaces for different perceptual attributes such as hardness, roughness, and friction, and they are used to define the physical space. In this paper, we question whether these state-of-the-art tactile features corresponding to hardness, roughness, and friction explain the overall dissimilarity/similarity rating between different texture pairs or not. Our approach follows two steps. Firstly, we investigate whether the regressor-based distance learned from these handcrafted features corresponds to perceptual space. Different regressors are learned on the fixed set of features. The results show that the regressor-based learning of distance between texture pairs attains a moderate level of agreement (in terms of mean square error, Spearman correlation, and Pearson correlation) with perceptual dissimilarities of corresponding pairs (i.e., perceptual space). Different regressors have provided almost similar accuracy. Second, we determine the relevance of a structure in tactile features in finding mapping with the perceptual space with the help of the permutation-test-based approach, where the features are perturbed. We hypothesize that if there is any structure in physical space, then any perturbation to it will provide a bad mapping function between the physical and perceptual spaces. The permutation test results reject the hypothesis, indicating that the employed tactile features (corresponding to hardness, roughness, and friction) are not optimal to justify the overall perceptual dissimilarity between texture pairs. The results of the study may be employed in different applications of haptics in the area of human-computer interaction, tele-robotics, augmented and virtual reality, and material classification.

  • Explainable Machine Learning: The importance of a system-centric perspective [Lecture Notes]

    by Manish Narwaria

    The landscape in the context of several signal processing applications and even education [1] appears to be significantly affected by the emergence of machine learning (ML) and particularly deep learning (DL). The main reason for this is the ability of DL to model complex and unknown relationships between signals and the tasks of interest. In particular, supervised DL algorithms have been fairly successful at recognizing perceptually or semantically useful signal information in different applications (e.g., identifying objects or regions of interest from image/video signals or recognizing spoken words from speech signals, such as in speech recognition). In all of these tasks, the training process uses labeled data to learn a mapping function (typically implicitly) from signals to the desired information (class label or target label). The trained DL model is then expected to correctly recognize/classify relevant information in a given test signal. Therefore, a DL-based framework is, in general, very appealing since the features and characteristics of the required mapping are learned almost exclusively from the data without resorting to explicit model/system development.