Sequence Recognition in Bharatnatyam dance
Himadri Bhuyan, Rohit Dhaipule, Partha Pratim Das
cs.CV · cs.LG
Quick verdict
Reports 98.7% Adavu sequence recognition via CNN+SVM+Edit Distance on Kinect data; solid engineering, but the headline accuracy rests on an oracle assumption.
Summary
This CVIP 2022 paper recognizes sequences of Bharatanatyam Adavus (13 classes, 51 variations) from Kinect V1 RGB video, combining key-posture recognition, motion recognition, and sequence matching. Frames are background-subtracted using depth masks; a small two-convolution, three-FC CNN classifies 182 key-posture classes at 99.46% accuracy, outperforming AlexNet, VGG16, and a ResNet baseline on this data. Motions are represented as HOG features of Motion History Images and classified by a one-vs-one SVM over 54 classes (filtered from 334 by a >=12-sample threshold) at roughly 84%. Adavu identity is decided by Edit-Distance matching of the recognized sequence against a ground-truth dictionary: 77.51% using postures alone over all 51 variations (100% on the Natta/Mettu subset of prior work), and 98.66% (1623/1645 sequences) when motions are included, with prediction 40-76x faster than the authors' earlier SVM pipeline. Main limitations: only 54 of 334 motion classes are modeled, and in the headline setting excluded motions are assumed correctly recognized; motion accuracy figures are internally inconsistent (84.08% vs 84.86% vs a recomputed 84.69%); evaluation uses a single split with no confidence intervals or significance tests; and Tatta-1/Tatta-8 are provably indistinguishable without audio. The annotated dataset is published online; code is not.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A compact CNN trained on background-subtracted frames recognizes 182 Bharatanatyam key-posture classes at 99.46% accuracy, outperforming AlexNet, VGG16, and ResNet on this dataset.
- HOG features of Motion History Images classified by a one-vs-one SVM recognize 54 motion classes at about 84% accuracy.
- Edit-Distance matching of recognized KP+motion sequences identifies all 51 Adavu variations at 98.66% accuracy (1623/1645 sequences), versus 77.51% using postures alone.
- The proposed pipeline predicts 40-76x faster than the authors' earlier SVM+Edit-Distance system [13] (0.316 s vs 12.88 s per sequence).
- Tatta-1 and Tatta-8 are indistinguishable from video because they share identical KP/motion sequences and differ only in the accompanying Sollukattu audio.
Strengths
First sequence-level treatment that includes motions, not just postures
Prior Adavu recognizers ([28], [21], [13], [15]) used KP sequences only; Approach-2 adds motion recognition and lifts accuracy from 77.51% (KP-only, 1645 sequences) to 98.66% (Sec 8.1, Table 6), a clearly scoped increment over the authors' own [13].
Evaluation scale well beyond prior 8-12 Adavus
Covers 13 Adavus / 51 variations / 1645 sequences vs one or two Adavus in [21], [13], [28]; per-class KP results with error analysis are published (Table 4, e.g., P94 at 57.14% misread as P95).
Efficiency analysis rare in this literature
Compact CNN competitive with deeper architectures on this data
Transparent error analysis and released data
Weaknesses
Headline accuracy embeds an oracle assumption
Internally inconsistent motion-recognition numbers
No uncertainty quantification
Potential leakage in sequence evaluation
Reporting sloppiness undermines trust
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What accuracy does this paper achieve in recognizing Bharatanatyam Adavus?
On 1645 test sequences covering 51 Adavu variations, the full pipeline (CNN key postures + SVM motions + Edit Distance matching) reaches 98.66% sequence-recognition accuracy; using key postures only it reaches 77.51%. Key postures alone are classified at 99.46% and motions at roughly 84%.
What dataset is used and is it public?
The authors use their own Kinect V1 recordings (RGB, depth, skeleton) of Bharatanatyam Adavus with frame-level annotations of key postures and motions — 13 Adavus and 51 variations after discarding erroneous data. The annotated dataset is published online via the IIT Kharagpur HCI lab page; the paper does not release source code.
Why can't the system distinguish Tatta-1 from Tatta-8?
Both variants contain the same key posture (P153) and the same alternating motions (M257, M258) in the same order; they differ only in the accompanying Sollukattu musical beats. The authors note the two can only be separated using audio, and the 22 errors in the headline result are exactly these confusions.
How does this work compare to earlier Adavu recognition methods?
Earlier systems (Sharma 2013; Kale and Patil 2015; Mallick et al. 2022; Bhuyan et al. 2021) used key postures only and covered 8-12 Adavus, with accuracies of about 80.55-99.38%. This paper adds motion recognition, evaluates on all 51 variations (1645 sequences), and reports roughly 40-76x faster prediction than its closest predecessor.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures (workflow, CNN architecture, misclassified poses, MHI examples) cannot be assessed from extracted text, and at least one figure reference appears mislabeled (MHI examples cited as Fig. 1).
- Tables were flattened by text extraction; per-class KP results (Table 4) were partially parsed and cross-checked arithmetically where possible.
- The train/test split protocol exists only in an external dataset page [3], so leakage risk could not be verified from the paper alone; the abstract in the supplied metadata is truncated mid-sentence but appears complete in the body text.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.