Foundation Models for 3D Scenes: DINOv2 vs. CLIP for Instance Differentiation

Table of Links

Abstract and 1 Introduction

1. Related Works

   2.1. Vision-and-Language Navigation

   2.2. Semantic Scene Understanding and Instance Segmentation

   2.3. 3D Scene Reconstruction

2. Methodology

   3.1. Data Collection

   3.2. Open-set Semantic Information from Images

   3.3. Creating the Open-set 3D Representation

   3.4. Language-Guided Navigation

3. Experiments

   4.1. Quantitative Evaluation

   4.2. Qualitative Results

4. Conclusion and Future Work, Disclosure statement, and References

2.2. Semantic Scene Understanding and Instance Segmentation

f 3D scenes. This domain has been thoroughly explored using closed-set vocabulary methods, including our prior work [1], which utilizes Mask2Former [7] for image segmentation. Various studies [18, 19, 20] have adopted a similar approach to achieve object segmentation, resulting in a closed-set framework. While these methods are effective, they are constrained by the limitation of predefined object categories. Our approach employs SAM [21] to acquire segmentation masks for open-set detection. Moreover, our methodology, distinct from many existing techniques that depend heavily on extensive pre-training or fine-tuning, integrates these models to forge a more comprehensive and adaptable 3D scene representation. This emphasizes enhanced semantic understanding and spatial awareness.

\ To improve the semantic understanding of the objects detected within our images, we harness detailed feature representations using two foundational models: CLIP [9] and DINOv2 [10]. DINOv2, a Vision Transformer trained through self-supervision, recognises pixel-level correspondences between images and captures spatial hierarchies. Compared to CLIP, DINOv2 more effectively distinguishes between two distinct instances of the same object type, which poses challenges for CLIP.

\ It’s crucial to differentiate individual instances following the semantic identification of objects. Early methods employed a Region Proposal Network (RPN) to predict bounding boxes for these instances [22]. Alternatively, some strategies suggest a generalized architecture for managing panoptic segmentation [23]. In our preceding approach, we utilized the segmentation model Mask2Former [7], which employs an attention mechanism to isolate object-centric features. Recent research also tackles semantic scene understanding using open vocabularies [24], utilizing multi-view fusion and 3D convolutions to derive dense features from an open-vocabulary embedding space for precise semantic segmentation. Our current pipeline leverages Grounding DINO [25] to generate bounding boxes, which are then input into the Segment Anything Model (SAM) [21] to produce individual object masks, thus enabling instance segmentation within the scene.

\

:::info Authors:

(1) Laksh Nanwani, International Institute of Information Technology, Hyderabad, India; this author contributed equally to this work;

(2) Kumaraditya Gupta, International Institute of Information Technology, Hyderabad, India;

(3) Aditya Mathur, International Institute of Information Technology, Hyderabad, India; this author contributed equally to this work.

(4) Swayam Agrawal, International Institute of Information Technology, Hyderabad, India;

(5) A.H. Abdul Hafez, Hasan Kalyoncu University, Sahinbey, Gaziantep, Turkey;

(6) K. Madhava Krishna, International Institute of Information Technology, Hyderabad, India.

:::

---

:::info This paper is available on arxiv under CC by-SA 4.0 Deed (Attribution-Sharealike 4.0 International) license.

:::

\