Towards Pixel-level VLM Perception via Simple Points Prediction

Introduction

[!Note] This is Qwen2.5-VL version of SimpleSeg, a dense architecture with 7B paramters.

We present SimpleSeg, a strikingly simple yet highly effective approach to endow Multimodal Large Language Models (MLLMs) with native pixel-level perception. Our method reframes segmentation as a simple sequence generation problem: the model directly predicts sequence of points (textual coordinates) delineating object boundaries, entirely within its language space. To achieve high fidelity, we introduce a two-stage SFT→RL training pipeline, where Reinforcement Learning with an IoU-based reward refines the point sequences to accurately match ground-truth contours. We find that the standard MLLM architecture possesses a strong, inherent capacity for low-level perception that can be unlocked without any specialized architecture. On segmentation benchmarks, SimpleSeg achieves performance that is comparable to, and often surpasses, methods relying on complex, task-specific designs. This work lays out that precise spatial understanding can emerge from simple point prediction, challenging the prevailing need for auxiliary components and paving the way for more unified and capable VLMs.

Method

In this work, we explore the limits of MLLM pixel-level perception by predicting the next point in a contour with the simplest approach possible. Without introducing any complex architectures or special patterns, we show how even minimalistic point prediction can achieve effective segmentation at the pixel level.

Key Benefits

• Simplicity: SimpleSeg requires no specialized modules and adheres to the standard MLLM architecture, it can be seamlessly and efficiently integrated as a new, core pre-training task for foundation models, similar to visual grounding.
• Task Generality: By framing segmentation as a text-generation problem, our approach is inherently flexible. The model can be easily adapted to a wide range of vision-language tasks that require precise spatial localization.
• Interpretable Output: The model generates explicit, human-readable coordinate sequences instead of dense pixel masks. This transparency simplifies debugging and makes the output directly usable for downstream applications like interactive editing or tool use.

Performance

• Referring Expression Segmentation results

• Referring Expression Comprehension results

Model Usage

Inference

We recommend using vLLM for production deployment. Requires vllm>=0.12.0 with --trust-remote-code.

First, start the vLLM server:

vllm serve sthui/SimpleSeg-Qwen2.5-VL \
  --trust-remote-code \
  --tensor-parallel-size 4 \
  --served-model-name SimpleSeg-Qwen2.5-VL \
  --host 0.0.0.0 \
  --port 8000

Then run the following code to inference:

import base64
from openai import OpenAI

# vLLM server configuration
VLLM_BASE_URL = "http://localhost:8000/v1"
MODEL_NAME = "SimpleSeg-Qwen2.5-VL"  # Should match --served-model-name in vllm serve

def encode_image(image_path: str) -> str:
    """Encode image to base64 string."""
    with open(image_path, "rb") as f:
        return base64.b64encode(f.read()).decode()

def inference(image_path: str, instruction: str) -> str:
    """Run GUI grounding inference via vLLM."""
    client = OpenAI(base_url=VLLM_BASE_URL, api_key="EMPTY")

    messages = [
        {
            "role": "user",
            "content": [
                {
                    "type": "image_url",
                    "image_url": {"url": f"data:image/png;base64,{encode_image(image_path)}"}
                },
                {"type": "text", "text": instruction},
            ],
        },
    ]

    response = client.chat.completions.create(
        model=MODEL_NAME,
        messages=messages,
        max_tokens=4096,
        temperature=0,
    )

    return response.choices[0].message.content

# Example usage
image_path = "./octopus.png"
instruction = "Output the polygon coordinates of octopus in the image."

response = inference(image_path, instruction)
print("Model output:", response)

Decode the polygons and masks from the response string

import re
import json
import pycocotools.mask as mask_utils

class RegexPatterns:
    BOXED_PATTERN = r'\\boxed\{([^}]*)\}'
    BLOCK_PATTERN = r'^```$\r?\n(.*?)\r?\n^```$'

    NON_NEGATIVE_FLOAT_PATTERN = (
        r'(?:[1-9]\d*\.\d+|0\.\d+|\d+)'
    )

    BBOX_PATTERN = rf'\[\s*({NON_NEGATIVE_FLOAT_PATTERN})\s*,\s*({NON_NEGATIVE_FLOAT_PATTERN})\s*,\s*({NON_NEGATIVE_FLOAT_PATTERN})\s*,\s*({NON_NEGATIVE_FLOAT_PATTERN})\s*\]'
    POINT_PATTERN = (
        rf'\[\s*({NON_NEGATIVE_FLOAT_PATTERN})\s*,\s*({NON_NEGATIVE_FLOAT_PATTERN})\s*\]'
    )

    POLYGON_PATTERN = rf'\[\s*{POINT_PATTERN}(?:\s*,\s*{POINT_PATTERN})*\s*\]'

polygon_matches = [
    m.group(0) for m in re.finditer(RegexPatterns.POLYGON_PATTERN, response, re.DOTALL)
]
pred_polygons = []
for polygon_match in polygon_matches:
    polygon = json.loads(polygon_match)
    pred_polygons.append(polygon)

pred_masks = []
for pred_polygon in pred_polygons:
    pred_polygon = np.array(pred_polygon) * np.array([width, height])
    rle = mask_utils.frPyObjects(pred_polygon.reshape((1, -1)).tolist(), height, width)
    mask = mask_utils.decode(rle)
    mask = np.sum(mask, axis=2, keepdims=True)
    pred_masks.append(mask)
pred_mask = np.sum(pred_masks, axis=0)
pred_mask = pred_mask.sum(axis=2)
pred_mask = (pred_mask > 0).astype(np.uint8)