tungitachi/vietnamese-traffic-sign-vqa-1k

Vietnamese Traffic Sign VQA — Balanced 1K Test Set

A small, deduplicated and balanced Visual Question Answering (VQA) test set for Vietnamese traffic signs. Every example is an image + a Vietnamese question + an answer. The set is intended purely for evaluation (all examples are test).

Quick start

from datasets import load_dataset

ds = load_dataset("tungitachi/vietnamese-traffic-sign-vqa-1k", split="test")
ex = ds[0]
ex["image"]          # PIL.Image
ex["question"]       # e.g. "Trong ảnh có biển ... không?"
ex["answer"]         # e.g. "Có"
ex["question_type"]  # one of 7 types

Contents

QA pairs	1,000
Images	622 (every image appears; 1–4 QA each)
Question types	7, balanced (~143 each)
Sign classes covered	47
Split	test only
Language	Vietnamese

Fields

• image — the traffic-scene image (datasets.Image)
• question / answer — Vietnamese question and its answer
• question_type — one of presence, count, sign_type, color, shape, location, attribute
• question_type_orig — original type before the yes_no+negative → presence merge
• answer_type — yes_no / number / other
• evidence_object_ids — object ids the answer refers to (0–4)
• image_id, question_id, source, label_source

Per-image object/bounding-box metadata (class, bbox_yolo, shape, colour, position) is provided separately in metadata/objects.jsonl.

How it was built

This is a derivative of Anakonkai/vietnamese-traffic-sign-vqa (rule-based auto-generated VQA over the Kaggle VNTS Vietnamese traffic-sign images). The source had two issues this set addresses:

1. Near-duplicate frames. The base images are video frames; many are near-identical. They were deduplicated with DINOv2 image embeddings (cosine ≥ 0.90 = duplicate), reducing 2,736 → 622 distinct images (a class-coverage safeguard keeps all 47 classes).
2. Answer/language-prior bias. In the source, a blind model could score ~55% by ignoring the image — notably yes_no was always “Có” and negative always “Không”. Here yes_no+negative are merged into one presence type with balanced Có/Không (so the answer depends on the image), and within every type the answers are balanced.

After balancing, the blind majority-answer baseline is ~22% (down from ~55%), so accuracy above this reflects actual visual understanding. Report results per question_type, since types differ in difficulty.

Blind-baseline by type (always guessing the majority answer)

presence 50% · shape 34% · count 26% · color 25% · location 11% · sign_type 6% · attribute 4%.

Intended use & limitations

• Evaluation only — there is no train split; do not train on it.
• Questions/answers are template-based (rule-generated), so phrasing variety is limited.
• Vietnamese only.

License & attribution

Released under CC BY-SA 4.0, inheriting the licence of the source data. Please credit the original Kaggle VNTS dataset and Anakonkai/vietnamese-traffic-sign-vqa.