DSI-Bench: A Benchmark for Dynamic Spatial Intelligence

Paper | Project Page | Code

Abstract

Reasoning about dynamic spatial relationships is essential, as both observers and objects often move simultaneously. Although vision-language models (VLMs) and visual expertise models excel in 2D tasks and static scenarios, their ability to fully understand dynamic 3D scenarios remains limited. We introduce Dynamic Spatial Intelligence and propose DSI-Bench, a benchmark with nearly 1,000 dynamic videos and over 1,700 manually annotated questions covering nine decoupled motion patterns of observers and objects. Spatially and temporally symmetric designs reduce biases and enable systematic evaluation of models' reasoning about self-motion and object motion. Our evaluation of 14 VLMs and expert models reveals key limitations: models often conflate observer and object motion, exhibit semantic biases, and fail to accurately infer relative relationships in dynamic scenarios. Our DSI-Bench provides valuable findings and insights about the future development of general and expertise models with dynamic spatial intelligence.

Sample Usage

This section provides a quick guide to set up the environment, download the dataset, perform inference, and evaluate model performance.

1. Install dependency

pip install -r requirements.txt

2. Download full dataset

huggingface-cli download --repo-type dataset Viglong/DSI-Bench --local-dir DSI-Bench

3. Inference with Qwen API

Here we provide a sample for testing on DSI-bench using the Qwen API.

import os
import re
import pandas as pd
from tqdm import tqdm
from concurrent.futures import ThreadPoolExecutor, as_completed
import dashscope

# Configuration
DASHSCOPE_API_KEY = "YOUR_DASHSCOPE_API_KEY_HERE"  # Replace or load from env

VLM_MODEL = "qwen2.5-vl-32b-instruct"   # e.g., "qwen2.5-vl-72b-instruct"
VIDEO_AUG = "std"                       # Video variant: 'std', 'hflip', etc.
NUM_WORKERS = 2
MAX_RETRIES = 10
FPS = 5                                 # Frames per second for video input

# Relative paths (relative to this script)
METADATA_BASE_DIR = "/path/to/metadatas"# Contains {VIDEO_AUG}.csv
VIDEO_BASE_DIR = "/path/to/videos"      # Structure: videos/{VIDEO_AUG}/xxx.mp4
OUTPUT_BASE_DIR = "/path/to/outputs"    # Output: outputs/{VIDEO_AUG}/{model}.csv

# Prompt
rawqa_prompt = """You are a vision-language expert.
You are given a clip of video and your task is to answer a question about the video.
You only need to provide *ONE* correct answer selecting from the options listed below. 
For example, if you think the correct answer is 'A' from 'A. Above B. Under C. Front D. Behind', 
your response should **only** be '<answer>A</answer>'.
Please answer the question in this format strictly:
<answer>[A, B, C, or D]</answer>
"""

def extract_single_choice_with_word_boundary(text):
    """
    Extract the answer letter (A/B/C/D) from <answer>X</answer> in the response.
    Returns None if not found.
    """
    match = re.search(r"<answer>\s*([A-D])\s*</answer>", text, re.IGNORECASE)
    return match.group(1).upper() if match else None

# Call VLM with video and question
def query_vlm(mp4_path, question, options):
    """
    Send video and question to Qwen-VL via DashScope API.
    Returns raw model response text.
    """
    question_and_options = f"
Question:
{question} {options}"
    messages = [
        {
            "role": "user",
            "content": [
                {"video": f"file://{mp4_path}", "fps": FPS},
                {"text": rawqa_prompt + question_and_options}
            ]
        }
    ]
    
    response = dashscope.MultiModalConversation.call(
        api_key=DASHSCOPE_API_KEY,
        model=VLM_MODEL,
        messages=messages,
        max_length=2048,
        stream=False,
        top_k=1
    )
    return response.output.choices[0].message.content[0]["text"]

# Process a single sample
def process_sample(meta, idx):
    """
    Process one video-question sample.
    Returns (raw_response, extracted_answer) or (None, None) on failure.
    """
    mp4_path = os.path.join(VIDEO_BASE_DIR, VIDEO_AUG, meta["relative_path"][idx])
    question = meta["question"][idx]
    options = meta["options"][idx]

    try:
        raw_response = query_vlm(mp4_path, question, options)
        final_answer = extract_single_choice_with_word_boundary(raw_response)
        return raw_response, final_answer
    except Exception as e:
        return None, None

# Batch processing with retries
def run_batch_inference(meta, num_workers=NUM_WORKERS, max_retries=MAX_RETRIES):
    """
    Run inference on all samples with parallel execution and retry logic.
    Ensures output order matches input order.
    """
    print(f"Running inference with model: {VLM_MODEL}, video variant: {VIDEO_AUG}")
    n = len(meta)
    results = [None] * n

    def worker(idx):
        raw, ans = process_sample(meta, idx)
        return idx, raw, ans

    def run_parallel(indices, desc="Processing"):
        temp = [None] * n
        with ThreadPoolExecutor(max_workers=num_workers) as executor:
            futures = {executor.submit(worker, i): i for i in indices}
            for future in tqdm(as_completed(futures), total=len(futures), desc=desc):
                idx, raw, ans = future.result()
                temp[idx] = {"result_text": raw, "final_answer": ans}
        return temp

    # Initial run
    current_results = run_parallel(list(range(n)), desc="Initial Run")
    for i, res in enumerate(current_results):
        results[i] = res

    # Retry failed samples
    for retry in range(1, max_retries + 1):
        failed = [i for i in range(n) if results[i]["final_answer"] is None]
        if not failed:
            print(f"All samples succeeded after {retry - 1} retries.")
            break

        print(f"Retry {retry}/{max_retries} for {len(failed)} failed samples...")
        retry_results = run_parallel(failed, desc=f"Retry {retry}")
        for i in failed:
            results[i] = retry_results[i]

    # Handle permanently failed samples
    for i in range(n):
        if results[i]["final_answer"] is None:
            results[i]["final_answer"] = "E"  # Default error answer
            if results[i]["result_text"] is None:
                results[i]["result_text"] = ""

    success_count = sum(1 for r in results if r["final_answer"] != "E")
    print(f"Completed. Success: {success_count}/{n}")
    return results

if __name__ == "__main__":
    for aug in ["std", "hflip", "reverse", "reverse_hflip"]:
        VIDEO_AUG = aug
        meta_path = os.path.join(METADATA_BASE_DIR, f"{VIDEO_AUG}.csv")
        df = pd.read_csv(meta_path)
        print(f"Loaded metadata: {meta_path} with {len(df)} samples")

        # Set directories
        output_dir = os.path.join(OUTPUT_BASE_DIR, VIDEO_AUG)
        os.makedirs(output_dir, exist_ok=True)

        # Run inference
        results = run_batch_inference(df)

        # Save results
        output_file = os.path.join(output_dir, f"{VLM_MODEL}.csv")
        pd.DataFrame(results).to_csv(output_file, index=False)
        print(f"Results saved to: {output_file}")

4. Evaluate model performance

Use the following code to get the Sample-wise Accuracy and Group-wise Accuracy of the model.

import pandas as pd
import os
from typing import Dict

# ==============================
# Configuration
# ==============================
VLM_MODEL = "qwen2.5-vl-32b-instruct"  # Model name
VIDEO_AUGS = ["std", "reverse", "hflip", "reverse_hflip"]

# Base paths (relative to project root)
META_BASE_PATH = "/path/to/metadatas"        # Contains {aug}.csv
OUTPUT_BASE_PATH = "/path/to/outputs"        # Contains {aug}/{VLM_MODEL}.csv

# Category name mapping
CATE_NAMES = [
    "Obj:static cam",
    "Obj:moving cam",
    "Cam:static scene",
    "Cam:dynamic scene",
    "Obj-Cam distance",
    "Obj-Cam orientation"
]

def load_all_data() -> Dict[str, Dict[str, pd.DataFrame]]:
    """
    Load metadata and prediction results for all augmentations.
    Returns: {aug: {'meta': DataFrame, 'result': DataFrame}}
    """
    data_dict = {}

    for aug in VIDEO_AUGS:
        meta_path = os.path.join(META_BASE_PATH, f"{aug}.csv")
        res_path = os.path.join(OUTPUT_BASE_PATH, aug, f"{VLM_MODEL}.csv")

        if not os.path.exists(meta_path):
            raise FileNotFoundError(f"Metadata not found: {meta_path}")
        if not os.path.exists(res_path):
            raise FileNotFoundError(f"Result file not found: {res_path}")

        meta = pd.read_csv(meta_path)
        result = pd.read_csv(res_path)

        if len(meta) != len(result):
            raise ValueError(f"Length mismatch in {aug}: meta={len(meta)}, result={len(result)}")
        if "GT" not in meta.columns:
            raise ValueError(f"'GT' column missing in metadata: {meta_path}")
        if "final_answer" not in result.columns:
            raise ValueError(f"'final_answer' column missing in results: {res_path}")

        data_dict[aug] = {"meta": meta, "result": result}

    # Ensure all augmentations have the same number of samples
    lengths = [len(data_dict[aug]["meta"]) for aug in VIDEO_AUGS]
    if len(set(lengths)) > 1:
        raise ValueError(f"Inconsistent sample counts: {dict(zip(VIDEO_AUGS, lengths))}")

    return data_dict

def sample_wise_evaluation():
    """
    Treat all samples across all augmentations as independent.
    Compute per-category and overall accuracy.
    """
    print("=== Method 1: Independent Samples ===")
    data_dict = load_all_data()

    records = []
    for aug in VIDEO_AUGS:
        meta = data_dict[aug]["meta"]
        res = data_dict[aug]["result"]
        for i in range(len(meta)):
            gt = str(meta.iloc[i]["GT"]).strip()
            pred = str(res.iloc[i]["final_answer"]).strip()
            pred_letter = pred[0] if pred else ""
            correct = int(gt == pred_letter)
            records.append({"cate": meta.iloc[i]["cate"], "correct": correct})

    df = pd.DataFrame(records)
    acc_by_cat = df.groupby("cate")["correct"].mean()
    overall_acc = df["correct"].mean()
    print_metrics(acc_by_cat, overall_acc)

def group_wise_evaluation(n: int):
    """
    For each original question (4 views), count how many augmented views are correct
    under their own ground truth. If >= n are correct, count as robustly correct.
    """
    print(f"=== Method 2: Ensemble Voting (n>={n}) ===")
    data_dict = load_all_data()
    num_samples = len(data_dict[VIDEO_AUGS[0]]["meta"])

    records = []
    total_robust_correct = 0

    for i in range(num_samples):
        correct_count = 0
        cate = None
        for aug in VIDEO_AUGS:
            meta = data_dict[aug]["meta"]
            res = data_dict[aug]["result"]
            gt = str(meta.iloc[i]["GT"]).strip()
            pred = str(res.iloc[i]["final_answer"]).strip()
            pred_letter = pred[0] if pred else ""
            if gt == pred_letter:
                correct_count += 1
            if cate is None:
                cate = meta.iloc[i]["cate"]

        is_robust_correct = int(correct_count >= n)
        records.append({"cate": cate, "correct": is_robust_correct})
        total_robust_correct += is_robust_correct

    df = pd.DataFrame(records)
    acc_by_cat = df.groupby("cate")["correct"].mean()
    overall_acc = total_robust_correct / num_samples
    print_metrics(acc_by_cat, overall_acc)

def single_evaluation(aug: str):
    """
    Evaluate performance on a single augmentation variant.
    """
    if aug not in VIDEO_AUGS:
        raise ValueError(f"Invalid augmentation: {aug}. Choose from {VIDEO_AUGS}")

    print(f"=== Method 3: Single View Evaluation ({aug}) ===")
    data_dict = load_all_data()
    meta = data_dict[aug]["meta"]
    res = data_dict[aug]["result"]

    correct_list = []
    for i in range(len(meta)):
        gt = str(meta.iloc[i]["GT"]).strip()
        pred = str(res.iloc[i]["final_answer"]).strip()
        pred_letter = pred[0] if pred else ""
        correct_list.append(int(gt == pred_letter))

    df = pd.DataFrame({"cate": meta["cate"], "correct": correct_list})
    acc_by_cat = df.groupby("cate")["correct"].mean()
    overall_acc = df["correct"].mean()
    print_metrics(acc_by_cat, overall_acc)

def print_metrics(acc_by_cat: pd.Series, overall_acc: float):
    """
    Print per-category and overall accuracy in a readable format.
    """
    for cat, ratio in acc_by_cat.items():
        name = CATE_NAMES[cat] if cat < len(CATE_NAMES) else f"Category {cat}"
        print('category = {0} {1:<20}  Acc = {2:.2%}'.format(cat, name, ratio))
    print(f'
Overall Acc = {overall_acc:.2%}
')

if __name__ == "__main__":
    print(f"Model: {VLM_MODEL}")

    sample_wise_evaluation()
    group_wise_evaluation(n=3)

Citation

If you find this repository useful for your research, please use the following BibTeX entry:

@misc{zhang2025dsibenchbenchmarkdynamicspatial,
      title={DSI-Bench: A Benchmark for Dynamic Spatial Intelligence}, 
      author={Ziang Zhang and Zehan Wang and Guanghao Zhang and Weilong Dai and Yan Xia and Ziang Yan and Minjie Hong and Zhou Zhao},
      year={2025},
      eprint={2510.18873},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2510.18873}, 
}