WADE: A Reasoning-Annotated Benchmark for Multi-Instance Floating-Waste Grounding with Compact Vision-Language Models

Md. Asaduzzaman Shuvo; Ahsan Farabi; Md. Abdul Ahad Minhaz; Mahedi Hasan; Israt Khandaker; Ibrahim Khalil Shanto; Muhammad Nomani Kabir

cs.CVAug 24, 2026

WADE: A Reasoning-Annotated Benchmark for Multi-Instance Floating-Waste Grounding with Compact Vision-Language Models

Authors: Md. Asaduzzaman Shuvo, Ahsan Farabi, Md. Abdul Ahad Minhaz, Mahedi Hasan, Israt Khandaker, Ibrahim Khalil Shanto, Muhammad Nomani Kabir

Organizations: United International University

Abstract

Floating waste in inland waterways threatens aquatic ecosystems and requires timely monitoring under cluttered, multi-object conditions. Existing aquatic-waste datasets provide limited geographic coverage, sparse multi-instance annotations, and little supervision beyond boxes and labels. Compact vision-language models (VLMs) therefore remain insufficiently evaluated for jointly localizing, classifying, counting, and explaining floating waste. We introduce WADE, a reasoning-annotated benchmark containing 2,167 images from rural Bangladesh, 13,608 bounding boxes, and ten waste categories. Each annotation is associated with class-level recognition rules covering visual cues, likely confusions, and discriminative features. We evaluate six VLMs under zero-shot, two-shot, reasoning-guided, and fine-tuned settings using detection, counting, and hallucination metrics. For resource-efficient adaptation, we jointly fine-tune Qwen3-VL-2B on boxes, labels, and reasoning chains using QLoRA. Fine-tuning increases recall from 0.0248 to 0.2339 and F1 from 0.0257 to 0.2163, while reducing image-level hallucination from 0.6836 to 0.0883. However, over three-quarters of instances remain undetected, establishing WADE as a challenging benchmark for dense floating-waste grounding with compact VLMs.

Figures & tables

Figure 1: Overview of the WADE task. Floating-waste instances can be small, crowded, partially occluded, and visually similar to surrounding vegetation. Given a field image, a compact vision-language model must ground each visible instance and provide its bounding box, category label, and visual reasoning.

Class	N	%	Class	N	%
Water hyacinth	2,137	15.7	Algal bloom	1,110	8.2
Plastic bottle	1,933	14.2	Solid waste	1,008	7.4
Organic waste	1,605	11.8	Fabric waste	976	7.2
Industrial waste	1,563	11.5	Wood debris	887	6.5
Polythene	1,525	11.2	Foam waste	864	6.3
Images: 2,167			Mean: 6.28 instances/image

Table 1: Distribution of WADE’s 13,608 annotated instances.

Figure 2: WADE annotation example. Every visible target instance receives a bounding box and category label and is linked to its class-level reasoning chain. Selected reasoning fields are shown for readability.

Figure 3: Overview of the WADE methodology. Field images are exhaustively annotated with bounding boxes, taxonomy labels, and class-level reasoning chains. Six vision-language models are evaluated under zero-shot, two-shot, reasoning-prompted, and QLoRA fine-tuning regimes. Predictions are assessed using automatic metrics, LLM judges, and blind human evaluation.

		Grounding						Hallucination		Counting Error
Regime	Model	$\mathbf{P_{.5}}\uparrow$	$\mathbf{R_{.5}}\uparrow$	$\mathbf{F1_{.5}}\uparrow$	$\mathbf{F1_{\mathrm{ag}}}\uparrow$	$\mathbf{R_{.75}}\uparrow$	$\mathbf{N_{p}/N_{g}}$	$\mathbf{CHAIR_{i}}\downarrow$	$\mathbf{CHAIR_{s}}\downarrow$	$\mathbf{MAE}\downarrow$	$\mathbf{RMSE}\downarrow$
Zero-shot	LLaMA-3.2-11B $\dagger$	.0070	.0066	.0068	.0123	.0027	.94	.7621	.9839	2.187	3.610
Zero-shot	InternVL2-2B $\dagger$	.0022	.0046	.0030	.0043	.0000	2.09	.8639	.9960	4.325	5.232
Zero-shot	Qwen3-VL-2B $\dagger$	.0267	.0248	.0257	.0347	.0107	.93	.6722	.6836	2.540	3.889
Zero-shot	LFM2.5-VL-1.6B $\dagger$	.0031	.0048	.0043	.0110	.0019	1.55	.8374	.5291	3.430	5.739
Zero-shot	GPT-4o-mini $\ddagger$	.0339	.0153	.0211	.0293	.0013	.45	.5817	.4679	1.954	3.597

Table 2: Grounding, hallucination, and counting performance on the WADE test set. Grounding metrics use class-aware matching unless marked class-agnostic (

F1_{\mathrm{ag}}

N_{p}/N_{g}

is the ratio between predicted and ground-truth instance counts. CHAIR measures unsupported predictions.

\dagger

denotes open-source and

\ddagger

commercial models. Best values are shown in bold.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Hyperparameter	Value
Base model	Qwen3-VL-2B-Instruct
Quantization	4-bit NF4
LoRA rank	16
LoRA $\alpha$	32
LoRA dropout	0.05
Epochs	3

Appendix

Table 3: QLoRA fine-tuning configuration.

Regime	Model	Box	Reason.	Halluc.
Zero-shot	InternVL2-2B	2.1	4.0	2.5
Zero-shot	LLaMA-3.2-11B	3.0	5.5	2.6
Zero-shot	Qwen3-VL-2B	3.6	6.5	3.8
Zero-shot	GPT-4o-mini	3.9	7.0	5.1
Zero-shot	Gemini-2.5-Flash	4.1	7.8	4.2
Reasoning	Qwen3-VL-2B	4.3	8.0	5.8

Appendix

Table 4: LLM-as-judge evaluation averaged over three judges.

Regime	Model	Box	Reason.	Halluc.
Zero-shot	InternVL2-2B	1.5	3.5	2.5
Zero-shot	LLaMA-3.2-11B	3.0	4.0	3.5
Zero-shot	Qwen3-VL-2B	4.0	5.5	3.5
Zero-shot	GPT-4o-mini	4.5	6.0	6.5
Zero-shot	Gemini-2.5-Flash	5.0	7.0	6.5
Reasoning	Qwen3-VL-2B	4.5	6.0	4.5

Appendix

Table 5: Blind human evaluation averaged over two independent raters.

Vision-Language Grounding

Vision-Language Grounding links words and phrases to the objects, regions, actions, or moments they describe in images, video, and scenes.

See all 418 papers

MARINER: A 3E-Driven Benchmark for Fine-Grained Perception and Complex Reasoning in Open-Water Environments
Apr 9, 2026Xingming Liao, Ning Chen, Muying Shu +5VLM Evaluation Visual Question Answering
Guangdong University of Technology, Guangzhou, China
MDWD: A Street-Level Dataset for Municipal Solid Waste Detection in Dense Urban Environments
Jul 31, 2026Andrea Filiberto Lucas, Mark Bugeja, Carl James Debono +1Object Detection
University of Malta
FloodReasonBench: Benchmarking VLM Reasoning Segmentation for Embodied Flood Response at the Edge
Aug 15, 2026Rajat Bhattacharjya, Yoomee Jung, Minwoo Kim +4Reasoning Segmentation VLM Evaluation
University of California, Irvine · Kookmin University, South Korea

Apr 9, 2026J/K move · Enter open · S save

MARINER: A 3E-Driven Benchmark for Fine-Grained Perception and Complex Reasoning in Open-Water Environments

Xingming Liao, Ning Chen, Muying Shu, Yunpeng Yin+4

Guangdong University of Technology, Guangzhou, China

VLM Evaluation Visual Question Answering Fine-Grained Image Classification Object Detection

Open paper AskPDF arXiv

Fine-grained visual understanding and high-level reasoning in real-world open-water environments remain under-explored due to the lack of dedicated benchmarks. We introduce MARINER, a comprehensive benchmark built under the novel Entity-Environment-Event (3E) paradigm. MARINER contains 16,629 multi-source maritime images with 63 fine-grained vessel categories, diverse adverse environments, and 5 typical dynamic maritime incidents, covering fine-grained classification, object detection, and visual question answering tasks. We conduct extensive evaluations on mainstream Multimodal Large language models (MLLMs) and establish baselines, revealing that even advanced models struggle with fine-grained discrimination and causal reasoning in complex marine scenes. As a dedicated maritime benchmark, MARINER fills the gap of realistic and cognitive-level evaluation for maritime multimodal understanding, and promotes future research on robust vision-language models for open-water applications. Appendix and supplementary materials are available at https://lxixim.github.io/MARINER.

Abstract

Vision-Language Grounding

Explore similar work

MARINER: A 3E-Driven Benchmark for Fine-Grained Perception and Complex Reasoning in Open-Water Environments

MDWD: A Street-Level Dataset for Municipal Solid Waste Detection in Dense Urban Environments

FloodReasonBench: Benchmarking VLM Reasoning Segmentation for Embodied Flood Response at the Edge