💻 computer science

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

This paper introduces WADE, a reasoning-annotated benchmark featuring 2,167 images of floating waste in rural Bangladesh with detailed visual rules, to evaluate and improve the performance of compact vision-language models in localizing, counting, and explaining multi-instance waste under challenging conditions.

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

Published 2026-08-25

📖 5 min read🧠 Deep dive

CC BY 4.0

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

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Rivers, ponds, and canals are the veins of many landscapes, carrying life and water through communities. Yet, these waterways often become choked with floating debris, a mix of plastic bottles, tangled fabric, and organic matter that threatens the health of the ecosystem. For decades, scientists have relied on manual patrols or broad satellite images to track this pollution, but these methods struggle to see the small, scattered pieces of trash hidden among reeds and reflections. In recent years, a new kind of artificial intelligence has emerged, capable of not just identifying objects in a picture but also describing them in words. These systems, known as vision-language models, promise to act as tireless observers that can count, locate, and explain what they see. However, while these models are excellent at spotting a single, clear object, they often stumble when faced with a messy, crowded scene where dozens of items overlap, blend into the background, or are only partially visible. The question remains: can a computer truly understand a cluttered river surface well enough to help clean it up?

A team of researchers from United International University set out to answer this by creating a new test specifically designed for this difficult task. They built a dataset called WADE, which stands for a benchmark for floating-waste grounding. The team collected 2,167 real-world photographs from rural waterways in Bangladesh, capturing the messy reality of ponds and canals during different seasons and times of day. In these images, waste does not sit neatly in a line; it floats in clusters, often half-submerged or tangled with water hyacinths and algae. The researchers meticulously marked every single piece of trash they could find, drawing a box around 13,608 individual items ranging from plastic bottles to foam and wood debris. What makes this dataset unique is that for every type of trash, they also wrote a set of rules explaining how to tell it apart from similar-looking things. For example, they noted how to distinguish a plastic bottle from a piece of foam or how to spot organic waste that looks like a natural algal bloom. This added layer of reasoning was intended to teach the computer not just what to look for, but how to think about what it sees.

The researchers then tested six different artificial intelligence models against this new challenge, ranging from small, efficient programs that could run on modest computers to massive, powerful systems used by major technology companies. They asked these models to look at the river photos and list every piece of trash they saw, providing a location for each item, its name, and a short explanation of why they chose that name. When the models were asked to do this without any prior training on these specific images, the results were sobering. Even the most advanced commercial systems struggled to find the trash accurately. They often guessed the right number of items but placed the location boxes in the wrong spots, or they confidently described objects that were not actually there. The models seemed to understand the concept of waste but failed to pin it down to the exact pixels in the image, a problem known as poor spatial grounding.

To see if they could improve the situation, the researchers tried a few different strategies. First, they showed the models a couple of examples of how to answer the question, hoping this would guide them. This did not help much; in fact, for some models, it made them more hesitant and less likely to find the trash. Next, they gave the models the written rules about how to tell different types of waste apart, hoping this knowledge would sharpen their focus. This did make the models more careful, reducing the number of fake objects they invented, but it also made them miss even more of the real trash. The models became so focused on being sure that they stopped looking for the difficult, hidden items.

The most significant change came when the researchers taught one of the smaller models directly using the new dataset. They adjusted the model's internal settings so it could learn from the thousands of examples of boxes, labels, and reasoning rules they had prepared. This training transformed the model's performance. It began to find far more of the actual trash, correctly locating nearly a quarter of all the items it was tested on, a massive jump from its previous near-zero success rate. It also stopped inventing fake objects almost entirely. Remarkably, this small, trained model became better at finding the trash than the largest, most expensive commercial systems that had not been trained on this specific problem.

Despite this success, the researchers emphasize that the problem is far from solved. Even with the training, the best model still missed more than three-quarters of the trash in the images. It struggled particularly with items that were very small, heavily hidden, or blending perfectly into the water and plants. The study reveals a clear gap: current artificial intelligence can often describe what a scene looks like in words, but it is still very poor at pointing exactly where those things are when the scene is messy and complex. The researchers conclude that while teaching these models with specific, real-world examples is a powerful step forward, we are still a long way from having a system that can reliably monitor and count floating waste in the wild. The challenge remains to help computers see the world with the same clarity and attention to detail that a human observer would need to clean up a river.

Technical Summary: WADE – A Reasoning-Annotated Benchmark for Multi-Instance Floating-Waste Grounding

1. Problem Definition

Floating waste in inland waterways poses a significant threat to aquatic ecosystems, yet automated monitoring remains challenging due to the complex nature of these environments. Existing datasets and computer vision approaches often struggle with real-world water-surface imagery where waste objects are small, partially submerged, heavily occluded, visually similar to vegetation, and unevenly distributed.

While Vision-Language Models (VLMs) offer a unified interface for recognition, spatial grounding, and textual explanation, they face specific limitations in this domain:

Multi-Instance Grounding: Unlike identifying a single salient object, monitoring requires localizing every visible instance in a cluttered scene.
Object Hallucination: Generative VLMs frequently predict objects or categories not present in the image.
Lack of Reasoning Supervision: Current benchmarks provide bounding boxes and labels but lack structured reasoning chains that explain why an object belongs to a specific category, which is crucial for distinguishing waste from confounding elements like algae or reflections.

2. Methodology

The WADE Dataset

The authors introduce WADE, a new benchmark designed for dense, multi-instance floating-waste grounding using compact VLMs.

Data Collection: 2,167 in-situ images captured from rural waterways (ponds, canals, rivers) across 390 locations in Bangladesh. The collection covers diverse conditions (winter, monsoon, varying light) without artificial placement of waste.
Annotations: The dataset contains 13,608 annotated instances across 10 categories (e.g., plastic bottles, water hyacinth, algal bloom, industrial waste).
Reasoning Schema: A key innovation is the inclusion of class-level reasoning chains. Instead of annotating a unique explanation for every box, the authors define a structured, scene-invariant reasoning chain for each category. These chains include:
- Primary visual cues.
- Likely confusions (e.g., distinguishing plastic from foam).
- Contrastive decision rules.
- Fallback evidence for partial visibility.
Split Strategy: The data is split 75/10/15 (train/val/test) at the water-body level, ensuring that no specific pond, canal, or river appears in more than one partition. This creates a rigorous location-disjoint evaluation.

Experimental Setup

The study evaluates six VLMs (four open-weight: LLaMA-3.2-11B-Vision, InternVL2-2B, Qwen3-VL-2B-Instruct, LFM2.5-VL-1.6B; and two commercial: GPT-4o-mini, Gemini-2.5-Flash) under four regimes:

Zero-Shot: Strict instruction-only inference.
Two-Shot: In-context learning with two fixed training examples.
Reasoning-Guided Prompting: Providing class-level recognition rules without image examples.
Reasoning-Supervised Fine-Tuning: Jointly fine-tuning Qwen3-VL-2B using QLoRA on the complete ground-truth JSON (bounding boxes, labels, and reasoning chains).

Evaluation Metrics

Performance is measured using:

Grounding: Class-aware Precision, Recall, and F1 at IoU $\ge$ 0.5 and 0.75.
Hallucination: Instance-level and Image-level CHAIR scores (measuring unsupported predictions).
Counting: Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE).
Qualitative Assessment: Blind human evaluation and LLM judges assessing box quality, reasoning quality, and hallucination control.

3. Key Results

Baseline Performance (Zero-Shot)

Existing VLMs perform poorly on dense floating-waste grounding without adaptation.

Low Recall: Even the strongest zero-shot model (Gemini-2.5-Flash) achieves an F1 of only 0.1174. Compact models like Qwen3-VL-2B score an F1 of 0.0257.
Localization vs. Counting: Models often generate a number of boxes similar to the ground truth (e.g., Qwen3-VL-2B produces 93% of the expected count) but fail to localize them correctly (Recall of only 2.48%).
Hallucination: Zero-shot models exhibit high hallucination rates; for instance, InternVL2-2B hallucinates in 99.6% of test images.

Impact of Prompting

Two-Shot: Generally fails to improve grounding; in some cases (e.g., Qwen3-VL-2B), performance degrades significantly (F1 drops to 0.0077).
Reasoning-Guided: Improves precision and reduces hallucinations but at the cost of recall. For Qwen3-VL-2B, recall drops from 0.0248 to 0.0137, indicating that providing discriminative rules makes the model overly conservative, missing many small or occluded objects.

Impact of Fine-Tuning

Fine-tuning Qwen3-VL-2B with QLoRA yields the most significant improvements:

Performance Gains: Recall increases from 0.0248 to 0.2339, and F1 rises from 0.0257 to 0.2163.
Hallucination Reduction: Image-level hallucination drops dramatically from 0.6836 to 0.0883.
Comparison: The fine-tuned 2B model outperforms the strongest zero-shot commercial system (Gemini-2.5-Flash) in precision, recall, and F1.
Limitation: Despite these gains, the model still misses over 75% of instances at IoU 0.5 and nearly 90% at IoU 0.75.

Qualitative Findings

Both LLM judges and human evaluators noted that models can generate plausible reasoning descriptions even when their bounding boxes are inaccurate. This suggests that fluent reasoning does not necessarily correlate with successful spatial grounding.

4. Key Contributions

WADE Dataset: A real-world, multi-instance benchmark with 2,167 images and 13,608 annotations from rural Bangladesh, featuring a location-disjoint split to prevent data leakage.
Reasoning Annotations: A structured schema associating every instance with a class-level reasoning chain (visual cues, confusions, rules), moving beyond simple box-and-label supervision.
Comprehensive Evaluation: A systematic study of six VLMs across zero-shot, few-shot, prompting, and fine-tuned regimes, highlighting the specific challenges of dense environmental grounding.
Empirical Evidence: Demonstration that joint domain-specific supervision (boxes + labels + reasoning) significantly improves compact VLM performance, yet the task remains largely unsolved.

5. Significance and Claims

The paper positions WADE as a challenging benchmark that exposes the gap between current VLM capabilities and the requirements for reliable environmental monitoring.

Modest Claims: The authors explicitly state that while fine-tuning improves performance, the task is "far from solved." The best model still fails to detect the majority of annotated objects, particularly small or occluded ones.
Role of Reasoning: The paper claims that reasoning annotations are valuable for supervision but cautions that the observed improvements in fine-tuning are due to the joint supervision of boxes, labels, and reasoning, not reasoning alone (as no box-and-label-only ablation was performed).
Future Direction: The results suggest that for specialized environmental perception, targeted supervision is more effective than simply increasing model scale. However, the authors emphasize that future work must address small-object representation, resolution-aware encoding, and the separation of counting accuracy from localization accuracy.

In conclusion, the paper argues that while compact VLMs can be adapted to this domain, current architectures struggle with the "exhaustive multi-instance grounding" required for dense, cluttered waterways, establishing WADE as a necessary step toward more robust environmental AI.