BAV-Classroom-VQA Dataset

• The paper introduces a dataset tailored for visual question answering on classroom video clips to assess student and teacher interactions.
• BAV-Classroom-VQA is a domain-specific resource capturing real classroom settings, diverse lecture formats, and distinct behavioral cues.
• It features ethical data collection, standardized annotations, and benchmarking with models like LLaMA2, LLaMA3, QWEN3, and NVILA using quantitative metrics.

Searching arXiv for the specified paper and closely related context. arXiv search query: (Vu et al., 30 Jul 2025) BAV-Classroom-VQA is a dataset introduced for Visual Question Answering on real classroom video data in the paper "Exploring the Application of Visual Question Answering (VQA) for Classroom Activity Monitoring" (Vu et al., 30 Jul 2025). It is designed to enable automatic monitoring of student and teacher behaviors, spatial arrangements, and interactions, with the stated goal of helping educators obtain objective, real-time insights into classroom engagement and activities. The dataset is derived from short clips extracted from lectures and seminars at the Banking Academy of Vietnam and is accompanied by a benchmarking study of several open-source VQA models, specifically LLaMA2, LLaMA3, QWEN3, and NVILA (Vu et al., 30 Jul 2025).

1. Research setting and intended scope

BAV-Classroom-VQA targets classroom behavior analysis from video, rather than generic scene understanding. Its scope includes student and teacher behaviors, spatial arrangements, and interactions observable in real classroom recordings. The source material consists of lectures and seminars at the Banking Academy of Vietnam, and the dataset covers a variety of teaching formats, classroom sizes, and interaction types (Vu et al., 30 Jul 2025).

All raw recordings were provided by the International School – Banking Academy of Vietnam, Hanoi campus. The original repository contains 13 lecture videos totaling 11 h 58 min. The classroom settings vary in room size, seating arrangements, camera angle, and lighting conditions. Reported metadata fields, where available, include video_id, room_id, recording_date, resolution such as 1920×1080, frame_rate with the note that it is unspecified in the paper, camera_angle with values front/side/back, and lighting_condition with values bright/normal (Vu et al., 30 Jul 2025).

This positioning matters methodologically because the dataset is explicitly tied to classroom analytics rather than to synthetic or web-scale VQA corpora. A plausible implication is that domain-specific visual cues such as seating geometry, lecturer position, and collective student activity are central to the benchmark’s difficulty profile.

2. Data collection and ethical constraints

The data collection procedure is described as a four-step workflow. First, signed consent was obtained from all lecturers and students appearing on camera. Second, recordings were selected if they satisfied high image resolution, clear lighting, minimal motion blur, full-class framing, and observable, distinct student behaviors. Third, the raw 13 videos were imported into a workstation and transcoded to a standard format, namely MP4 with H.264. Fourth, each long recording was segmented into 20–30 s clips in order to isolate coherent activity windows, reduce annotator workload, and maintain behavioral context (Vu et al., 30 Jul 2025).

The ethical framework is stated explicitly. Permissions were secured from both the institution and the individuals appearing in the recordings, and full anonymization of personal identifiers was applied in published data. Data use is restricted to academic research under institutional review board guidelines of the Banking Academy of Vietnam (Vu et al., 30 Jul 2025).

These constraints are not incidental. They define the dataset as a controlled academic resource rather than an unrestricted public benchmark. A plausible implication is that downstream reproducibility depends not only on model code and evaluation scripts but also on access governance and institutional compliance.

3. Annotation pipeline, question taxonomy, and answer constraints

The annotation pipeline uses a custom web interface that presents each video clip alongside fields for question and answer. Annotation was performed by a team of 5 trained graduate-level students in educational technology. Their preparation consisted of a 2 full-day workshop covering VQA taxonomy, guidance on phrasing consistent question templates, and procedures for applying word-limit constraints. After initial annotation, each clip’s Q&A set was cross-reviewed by two other annotators, and disagreements were resolved by consensus (Vu et al., 30 Jul 2025).

The reported question categories, adapted from VideoQA literature, are: Factual, Spatial, Counting, Action Recognition, and Reasoning/Causal. The dataset supports both Close-Ended and Open-Ended formats. Close-Ended questions include Yes/No and Single-Choice with predefined options. Open-Ended questions use free-form answers constrained by guided prompts such as “Not more than 3 words” and “Only one word or number.” Excluded question types are Attribute classification, Comparative, and unconstrained Multiple-Choice (Vu et al., 30 Jul 2025).

The annotation guidelines impose two notable restrictions. Every Q&A pair must refer exclusively to the video frames within the clip, and word limits are strictly enforced to maintain consistency in evaluation. Inter-annotator agreement was monitored informally via cross-review, but exact $\kappa$ scores were not reported and are planned for future releases (Vu et al., 30 Jul 2025).

This schema makes answer brevity part of the benchmark definition rather than a post hoc evaluation preference. That design choice is particularly consequential for open-ended scoring, because it constrains semantic variation at the annotation stage before metrics such as ROUGE-L and BERTScore are applied.

4. Composition, notation, and data representation

The paper does not report exact counts of clips or Q&A pairs. Instead, the dataset specification introduces notation for future reference. Let $N$ denote the total number of QA pairs, which is not specified; let $V$ denote the total number of video clips, with an estimate of approximately $13$ videos $$\times \approx (11\text{h}58\text{m} \div 25\text{ s}) \approx 1{,}720$$ clips; and let $K$ denote the number of behavior categories, for example Factual, Spatial, Counting, Action, and Reasoning (Vu et al., 30 Jul 2025).

The class-distribution vector is represented as

$$\mathbf{c}=\bigl(c_{Factual},\,c_{Spatial},\,c_{Counting},\,c_{Action},\,c_{Reasoning}\bigr),$$

where $c_i$ is the fraction of QA pairs in category $i$. Suggested data splits are also given, with the explicit caveat that they are not yet published:

$$\text{train} : \text{val} : \text{test} = 0.7 : 0.15 : 0.15.$$

(Vu et al., 30 Jul 2025)

The folder structure is illustrated as BAV-Classroom-VQA/ with subdirectories videos/, annotations/, and a metadata.csv file. The annotation schema is exemplified through a JSON entry containing fields for video_id, question_id, question, options, answer, answer_type, format, and word_limit (Vu et al., 30 Jul 2025).

A representative entry is reported in the following form:

$$\times \approx (11\text{h}58\text{m} \div 25\text{ s}) \approx 1{,}720$$3

At the time of publication, access is available upon request from the corresponding author at [email protected] under a non-commercial, academic-use license. A public download link and a DOI are planned for the 2026 release (Vu et al., 30 Jul 2025).

A common misconception would be to treat the estimated clip count or the suggested split ratio as finalized dataset statistics. The specification explicitly distinguishes these from officially published counts.

5. Benchmark protocol and reported model performance

The benchmark distinguishes Closed-Ended and Open-Ended evaluation. For Closed-Ended questions, the reported metric is accuracy over the set of close-ended questions, with $N$0 denoting the number of close-ended questions, $N$1 the model answer, and $N$2 the ground truth (Vu et al., 30 Jul 2025). For Open-Ended questions, the reported metrics are ROUGE-L and BERTScore:

$N$3

and

$N$4

with

$N$5

and

$N$6

(Vu et al., 30 Jul 2025)

The reported closed-ended baseline results are as follows:

Model	Accuracy summary	Avg. response time
LLaMA2	Yes/No 80%; Single-Choice 61.9%	~20 s
LLaMA3	Yes/No 75%; Single-Choice 57.1%	~10 s
QWEN3	Yes/No 80%; Single-Choice 47.6%	5–10 s
NVILA	Yes/No 50%; Single-Choice 57.1%	30–60 s

For open-ended evaluation, the overall model scores are:

Model	BERTScore	ROUGE-L
NVILA	0.7462	0.4134
QWEN3	0.7353	0.4013
LLaMA3	0.7120	0.3593
LLaMA2	0.6877	0.3075

The granular scores by question type are also reported. For Factual questions, NVILA reaches $N$7 in BERTScore/ROUGE-L, QWEN3 $N$8, LLaMA3 $N$9, and LLaMA2 $V$0. For Spatial questions, the corresponding values are NVILA $V$1, QWEN3 $V$2, LLaMA3 $V$3, and LLaMA2 $V$4. For Counting questions, they are NVILA $V$5, QWEN3 $V$6, LLaMA3 $V$7, and LLaMA2 $V$8. For Object Det., they are NVILA $V$9, QWEN3 $13$0, LLaMA3 $13$1, and LLaMA2 $13$2. For Action Rec., they are NVILA $13$3, QWEN3 $13$4, LLaMA3 $13$5, and LLaMA2 $13$6. For Reasoning, they are NVILA $13$7, QWEN3 $13$8, LLaMA3 $13$9, and LLaMA2 $$\times \approx (11\text{h}58\text{m} \div 25\text{ s}) \approx 1{,}720$$0 (Vu et al., 30 Jul 2025).

The performance pattern suggests that the benchmark differentiates relatively well between question classes. Factual and some object-related prompts produce substantially higher scores than Action Recognition and Reasoning, which is consistent with the greater inferential load of behavior-centered classroom interpretation. That reading, however, remains an interpretation of the reported results rather than an explicit claim of causal difficulty in the source.

6. Recommended usage, extensibility, and current limitations

The dataset specification provides several recommendations for new experiments. It advises fine-tuning or instruction-tuning VQA models on BAV-Classroom-VQA to adapt to domain-specific cues such as classroom layouts and student attire. It also recommends maintaining open-ended answer constraints, especially word limits, during training so that model outputs remain consistent with evaluation metrics. In addition, any train/val/test split should balance the mix of question types so as to reflect the original distribution $$\times \approx (11\text{h}58\text{m} \div 25\text{ s}) \approx 1{,}720$$1 (Vu et al., 30 Jul 2025).

The proposed extensions are explicit: expansion to additional schools and camera setups for greater environmental diversity; introduction of temporal reasoning questions such as “What happened just before the student stood up?”; incorporation of finer-grained behavior labels such as note-taking versus reading; addition of multi-turn dialogues for interactive questioning; and evaluation on closed-campus deployments under real-time constraints for in-lecture assistance (Vu et al., 30 Jul 2025).

Several limitations are equally explicit. Exact counts of clips and QA pairs are not reported. Inter-annotator agreement was monitored informally, but exact $$\times \approx (11\text{h}58\text{m} \div 25\text{ s}) \approx 1{,}720$$2 scores were not reported. Access is controlled and non-commercial at publication time, with a public download link and DOI planned for 2026 rather than already available (Vu et al., 30 Jul 2025).

Taken together, these features place BAV-Classroom-VQA in a transitional category between a proof-of-concept benchmark and a fully standardized community resource. Its value lies in grounding VQA evaluation in real classroom recordings with defined annotation constraints and reported baselines, while its current incompleteness in released statistics and agreement reporting means that comparisons and replications should be interpreted with corresponding caution.