Understanding Autonomous Driving Datasets by Describing Differences between Image Subsets in Natural Language

AI summary

Overview

Research area: Computer vision and autonomous driving, specifically dataset introspection using Vision Language Models (VLMs) and a task called set difference captioning.

Technical level: Intermediate. The reader benefits from familiarity with object detection, image embeddings, and basic VLM concepts, but the paper's core idea — "describe in words how these two groups of images differ" — is explained concretely.

Scope: The paper adapts set difference captioning to autonomous driving by comparing object-centric image patches, introduces the AD-Diff Bench benchmark with web-scraped, annotation-filtered, and CLIP-filtered splits, and studies how existing methods behave on noisy (low purity) and sparse (low concentration) differences.

What This Paper Is About

Large autonomous driving datasets are hard to understand: existing analysis mostly relies on metadata, fixed label taxonomies, or manual inspection, which either scales poorly or only describes differences the label schema already anticipated. The paper's goal is to generate free-form natural-language descriptions of how one subset of driving images differs from another, so that developers can detect domain shift, gaps, and biases that predefined labels miss. To make this work on complex traffic scenes, the authors reformulate the task around objects rather than whole images, and they build a benchmark to measure how well this works in the driving domain.

Key Contributions

1. Object-centric set difference captioning. The authors reformulate set difference captioning for autonomous driving by extracting image patches centered on detected objects, which simplifies aggregation across large datasets and allows difference descriptions to be attributed to specific object instances or categories.
2. AD-Diff Bench. A new benchmark and evaluation protocol for object-centric set difference captioning in the driving domain, consisting of three splits — web-scraped, annotation-filtered, and CLIP-filtered — with ground-truth difference descriptions.
3. A first study of high-sparsity differences. The paper introduces dilution and the concentration parameter to AD-Diff Bench to test whether set difference captioning methods can detect differences that hold for only a tiny fraction of images in each set.
4. A single-stage baseline and open-weight methodology. The authors compare a two-stage proposer-ranker pipeline against a single-stage variant, restricting all experiments to open-weight models for reproducibility and deployment, and release implementation and benchmark at https://github.com/KIT-MRT/AD-Diff.

Main Findings

• Driving data is harder than web imagery. Accuracy is lowest on the annotation-filtered split and highest on the web-scraped split for most approaches. The authors attribute this to lower image-patch resolution, image noise, motion blur, underexposure, and possibly a larger domain shift from the VLMs' primarily web-scraped training data. For the two-stage image-based approach, Acc@1 on the web-scraped split was 0.73, versus 0.56 on annotation-filtered-60 and 0.64 on CLIP-filtered.
• The two-stage pipeline is more robust to sparsity. The two-stage image-based approach clearly outperformed the single-stage image-based approach on the web-scraped split (Acc@1 0.73 vs 0.64) and the CLIP-filtered split (0.64 vs 0.49), but performed similarly on annotation-filtered-60 (0.56 vs 0.53). The authors attribute this to the ranker's ability to aggregate sparse differences across the complete dataset.
• Image-based and caption-based proposers are comparable; the feature-based proposer lags. Across all three splits, image-based and caption-based proposers scored about the same — for example, on annotation-filtered-60, caption-based reached Acc@1 0.60 versus 0.56 for image-based — while the feature-based proposer was far behind (0.20 on annotation-filtered-60, 0.33 on web-scraped).
• Web-split difficulty labels are validated. The image-based two-stage approach scored Acc@1 of 0.84 on the easy sub-split, 0.75 on medium, and 0.59 on hard.
• Concentration: accuracy collapses below roughly 0.5. In dilution experiments on the annotation-filtered-39 split, accuracy remained stable until a concentration of about 0.5 (every second image from the dilution set) and then dropped rapidly. None of the tested approaches reliably described set differences at very low concentrations.
• Purity behaves similarly, except for the single-stage variant. Accuracy was fairly stable up to a purity of 0.5 and then approached zero at low purity. The single-stage proposer was the exception: even at purity 0 it correctly predicted the set difference in 35% of cases — which the authors argue is a false positive, not an advantage, because the ranker averages out cluster-level cues.
• Concentration is proposed as the better real-world metric. The authors argue that accuracy at low concentration cannot be "cheated" as easily as accuracy at low purity.
• The judge is largely reliable. Using gpt-oss-120b as an automatic judge, manual labeling of over 1000 hypotheses on VisDiffBench showed gpt-oss assigned the same score as the human annotator for 80.3% of hypotheses, with a mean absolute error of 0.104.
• An application example produced validatable insights. Comparing nuImages pedestrian patches from Singapore Queenstown against Boston Seaport, all top-8 hypotheses were validated with dataset annotations and external statistics. The highest-ranked was "people wearing hard hats" (AUROC 0.602), supported by 12.4% of persons labeled construction_worker in nuImages for Singapore versus 3.4% for Boston.

Methodology in Plain English

The task, set difference captioning, takes a target set A and a reference set B of images and asks for a natural-language description of a property that is more true of A than of B. The reference set is what makes the description meaningful: without it, descriptions are either trivial or endlessly exhaustive.

To adapt this to driving, the authors do not compare whole camera frames. Instead they extract an image patch around each object, using pre-trained 2D detectors or existing bounding box annotations. Each patch is enlarged by 50% from the raw bounding box to include context, and the object of interest is drawn with a red bounding box so the model can find it in crowded scenes. Patches are then grouped into subsets using dataset annotations or metadata, and compared.

The comparison pipeline has two stages. A proposer samples subsets from A and B and generates candidate difference hypotheses. The authors test three proposer styles — feeding images directly to the VLM (image-based), first captioning images and reasoning over the text (caption-based), or computing mean embeddings and using feature arithmetic (feature-based). All three use Qwen3-VL-30B-A3B-Instruct, with three generation rounds that each sample 20 images per set and produce ten hypotheses. A ranker then scores each hypothesis against the full datasets: it computes the cosine similarity between every image embedding and the hypothesis text embedding, treats that as a binary classifier score, and uses the resulting AUROC as the hypothesis score. Ranking uses SigLIP 2 Giant. A single-stage variant instead hands 100 images per dataset to the VLM in one pass and asks it to propose and rank differences together.

Evaluation uses gpt-oss-120b as an LLM judge, scoring each hypothesis as 0 (no match), 0.5 (partial match), or 1 (perfect match) against the ground truth, then averaging the best score in the top-N ranked hypotheses to get Acc@N.

For realism, the benchmark adds two knobs. Purity measures set cleanliness: values below 1 mean some images have been swapped between the two sets, and 0 means fully shuffled. Concentration measures sparsity: diluting a set of n_S images with n_D images from a third set gives c = n_S / (n_S + n_D), so low concentration means the true difference appears in only a tiny fraction of images. The authors illustrate why this matters for driving: of the 784,444 nuImages annotations for the class vehicle, only 42 are the sub-class ambulance, a concentration of roughly 0.000054.

Why This Matters

Impact on research. Dataset analysis today is bottlenecked by predefined labels: you can only find differences you already had a category for. Natural-language difference descriptions are open-ended and relative, so they can surface appearance, infrastructure, and environmental shifts outside the label taxonomy — the kind of subtle, safety-relevant shifts that could degrade a perception system after deployment. The paper also provides the first domain-specific benchmark for this task and shows that existing general-purpose benchmarks are not representative of driving data.

Real-world applications:

• Operational design domain (ODD) expansion. Before deploying a fleet in a new city, compare pedestrian (or vehicle) patches from the new location against the existing training distribution — as demonstrated with Singapore versus Boston.
• Safety diagnostics. The paper frames the question "Which conditions contribute to accidents/close-calls?" as a subset definition that can then be fed into set difference captioning.
• Rare-object auditing. Checking whether a dataset contains sufficient examples of safety-critical long-tail objects, illustrated by the ambulance example, where a sub-class represents an extremely small share of all vehicle annotations.
• Large-scale data curation in "data engines." Automating part of the record-label-retrain-deploy loop while keeping a human-in-the-loop interface for safety-relevant decisions.

Industry relevance. The restriction to open-weight models is deliberate: autonomous driving workflows often require data privacy and offline operation, which makes proprietary APIs impractical. The authors also note the single-stage approach can be orders of magnitude faster for large datasets, making it attractive when the goal is rapid hypothesis sampling rather than precise ranking.

Future Directions

• Closing the low-concentration gap. No tested approach reliably described set differences at very low concentrations, yet the paper's own ambulance example (concentration roughly 0.000054) shows that such rare differences are exactly what matters for safety.
• Making patch-level comparison competitive with whole-image context. Some differences are only identifiable with additional context, which is why patches are enlarged by 50%; the paper reports no ablation isolating how much this expansion contributes.
• Evaluating the object-centric formulation against user needs. The paper positions the tool as human-in-the-loop dataset introspection, but no user study of whether practitioners find the generated descriptions actionable is reported.
• Extending beyond camera images. The paper studies camera image patches only; whether the same pipeline should be adapted to other sensor modalities is not addressed.
• Diagnosing why driving data is harder. The authors leave open whether the performance gap stems from image resolution and quality or from genuine domain shift in VLM training data, a question they state cannot be resolved by the reported experiments.

Target Audience

Researchers and engineers working on autonomous driving datasets, data curation, and validation, who need scalable ways to characterize distribution shift and long-tail coverage. It is also relevant to practitioners applying vision language models to domain-specific industrial data, and to anyone building dataset introspection tooling, since the paper is explicit about which design choices (two-stage versus single-stage, image-based versus caption-based proposers, purity versus concentration evaluation) hold up under realistic sparsity. Readers looking for a first introduction to VLMs may find the proposer/ranker machinery dense, but the object-centric reformulation and the dilution experiments are accessible without deep background.

Authors’ abstract

Understanding the composition of large-scale autonomous driving datasets is essential for safety, robustness, and reliable operation across domains. For example, domain shift between locations could lead to the operating environment being misaligned with the training data, resulting in potentially dangerous performance degradation. Yet, existing data analysis pipelines largely rely on metadata, predefined labels, or manual inspection, which provide limited semantic insight or do not scale. This paper studies set difference captioning: given two subsets of images, the goal is to produce a natural-language hypothesis describing differences between the target and reference set. Building on a two-stage formulation, we adapt the method to autonomous driving by focusing on object-centric patches derived from object detection, which simplifies aggregation and enables attribution of differences to specific object instances or categories. To evaluate this setting in-domain, we introduce a new benchmark, AD-Diff Bench. Low-concentration experiments assess the suitability of set-difference-captioning approaches to sparse, real-world differences. We restrict our experiments to open-weight models to support reproducibility and ease of deployment. The proposed benchmark and analysis provide a step towards practical, human-interpretable dataset introspection for autonomous driving datasets. Our implementation and benchmark dataset are available at https://github.com/KIT-MRT/AD-Diff