--- license: cc-by-4.0 pipeline_tag: image-text-to-text library_name: transformers --- # LEGO: A Model for Multi-View 3D Scene Understanding This repository contains the official weights for **LEGO**, a baseline method for multi-view reasoning in 3D scene understanding. LEGO leverages knowledge from pre-trained 2D LVLMs (specifically fine-tuning a Fuyu-8B model) and is trained using the **TripAlign** pre-training dataset. It is evaluated on **MV-ScanQA**, a novel 3D question answering dataset designed to rigorously test multi-view compositional reasoning. LEGO achieves state-of-the-art performance on MV-ScanQA, as well as on existing benchmarks for 3D dense captioning and question answering. This model was presented in the paper [Advancing 3D Scene Understanding with MV-ScanQA Multi-View Reasoning Evaluation and TripAlign Pre-training Dataset](https://huggingface.co/papers/2508.11058). - 🏠 [Project Page](https://matthewdm0816.github.io/tripalign-mvscanqa) - 💻 [GitHub Repository](https://github.com/matthewdm0816/MV-ScanQA-TripAlign)
LEGO Teaser Image
## Overview of LEGO, MV-ScanQA, and TripAlign The **MV-ScanQA** dataset addresses limitations in existing 3D vision-language datasets by introducing questions that explicitly require integrating information from multiple views, thus rigorously testing multi-view compositional reasoning over distant objects. To facilitate training for such demanding scenarios, the **TripAlign** dataset is introduced. This large-scale, low-cost 2D-3D-language pre-training corpus contains 1M `<2D view, set of 3D objects, text>` triplets, providing richer, view-grounded multi-object multimodal alignment signals than previous single-object annotations. **LEGO** (Large-scale Multi-View Grounding Objective) is the baseline method developed to tackle the multi-view reasoning challenge in MV-ScanQA. It transfers knowledge from pre-trained 2D LVLMs (like Fuyu-8B, which this model fine-tunes) to the 3D domain with TripAlign. ## Usage This model is a PEFT (Parameter-Efficient Fine-Tuning) LoRA adapter built on top of the `adept/fuyu-8b` base model. You can load and use it with the `transformers` and `peft` libraries. First, ensure you have the necessary libraries installed: ```bash pip install transformers accelerate peft torch torchvision pillow ``` Below is a sample code for inference. Please note that the image pre-processing functions (`build_transform`, `find_closest_aspect_ratio`, `dynamic_preprocess`, `load_image`) are adapted from the original repository's usage patterns for Fuyu-based models. ```python import numpy as np import torch import torchvision.transforms as T from PIL import Image from torchvision.transforms.functional import InterpolationMode from transformers import AutoModelForCausalLM, AutoTokenizer from peft import PeftModel, PeftConfig IMAGENET_MEAN = (0.485, 0.456, 0.406) IMAGENET_STD = (0.229, 0.224, 0.225) def build_transform(input_size): MEAN, STD = IMAGENET_MEAN, IMAGENET_STD transform = T.Compose([ T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img), T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC), T.ToTensor(), T.Normalize(mean=MEAN, std=STD) ]) return transform def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size): best_ratio_diff = float('inf') best_ratio = (1, 1) area = width * height for ratio in target_ratios: target_aspect_ratio = ratio[0] / ratio[1] ratio_diff = abs(aspect_ratio - target_aspect_ratio) if ratio_diff < best_ratio_diff: best_ratio_diff = ratio_diff best_ratio = ratio elif ratio_diff == best_ratio_diff: if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]: best_ratio = ratio return best_ratio def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbnail=False): orig_width, orig_height = image.size aspect_ratio = orig_width / orig_height target_ratios = set( (i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if i * j <= max_num and i * j >= min_num) target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1]) target_aspect_ratio = find_closest_aspect_ratio( aspect_ratio, target_ratios, orig_width, orig_height, image_size) target_width = image_size * target_aspect_ratio[0] target_height = image_size * target_aspect_ratio[1] blocks = target_aspect_ratio[0] * target_aspect_ratio[1] resized_img = image.resize((target_width, target_height)) processed_images = [] for i in range(blocks): box = ( (i % (target_width // image_size)) * image_size, (i // (target_width // image_size)) * image_size, ((i % (target_width // image_size)) + 1) * image_size, ((i // (target_width // image_size)) + 1) * image_size ) split_img = resized_img.crop(box) processed_images.append(split_img) assert len(processed_images) == blocks if use_thumbnail and len(processed_images) != 1: thumbnail_img = image.resize((image_size, image_size)) processed_images.append(thumbnail_img) return processed_images def load_image(image_file, input_size=448, max_num=12): image = Image.open(image_file).convert('RGB') transform = build_transform(input_size=input_size) images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num) pixel_values = [transform(image) for image in images] pixel_values = torch.stack(pixel_values) return pixel_values # Define the base model and the LoRA adapter ID base_model_name_or_path = "adept/fuyu-8b" # Replace 'your-org/your-repo' with the actual model ID on Hugging Face Hub peft_model_id = "your-org/your-repo" # e.g., kmichiru/LEGO # Load the base model print(f"Loading base model: {base_model_name_or_path}...") base_model = AutoModelForCausalLM.from_pretrained( base_model_name_or_path, torch_dtype=torch.bfloat16, low_cpu_mem_usage=True, trust_remote_code=True, device_map="auto" # Use 'auto' to load across available devices ) tokenizer = AutoTokenizer.from_pretrained(base_model_name_or_path, trust_remote_code=True, use_fast=False) # Load the PEFT adapter weights on top of the base model print(f"Loading LoRA adapter: {peft_model_id}...") model = PeftModel.from_pretrained(base_model, peft_model_id).eval() print("Model loaded successfully!") # Example usage (replace with your image path and question) # You might need to download a sample image, e.g., from the GitHub repo # A dummy image for testing: # from PIL import ImageDraw # dummy_image = Image.new('RGB', (800, 600), color = 'red') # draw = ImageDraw.Draw(dummy_image) # draw.text((10,10), "Sample Image", fill=(0,0,0)) # dummy_image.save("sample_image.png") image_path = "sample_image.png" # Replace with path to a real image if not Path(image_path).exists(): print(f"Warning: Image '{image_path}' not found. Please provide a valid image path or create a dummy image.") # Exit or handle gracefully if no image is available for execution exit() pixel_values = load_image(image_path, max_num=6).to(torch.bfloat16).cuda() # Ensure image is on GPU generation_config = dict(max_new_tokens=1024, do_sample=True) question = "Describe the main objects in this 3D scene." # Example question # For a Fuyu model, the prompt format might be specific. Refer to Fuyu documentation. # This example uses a basic chat format. response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=None, return_history=True) print(f'User: {question} Assistant: {response}') # Example for 3D question answering (assuming the model outputs bounding box coordinates) question_with_bbox = "What is the bounding box of the chair in this scene?" response_bbox, history_bbox = model.chat(tokenizer, pixel_values, question_with_bbox, generation_config, history=None, return_history=True) print(f'User: {question_with_bbox} Assistant: {response_bbox}') ``` ## Citation If you find this codebase useful, please consider citing our work: ```bibtex @inproceedings{mo2025mvscanqa, title={Advancing 3D Scene Understanding with MV-ScanQA Multi-View Reasoning Evaluation and TripAlign Pre-training Dataset}, author={Mo, Wentao and Chen, QingChao and Peng, Yuxin and Huang, Siyuan and Liu, Yang}, booktitle={Proceedings of the 33rd ACM International Conference on Multimedia}, year={2025}, } ``` ## License This code repository and datasets are licensed under a [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/) license. Copyright (c) 2025 Wentao Mo.