Learning cross-modal occupancy knowledge: RadOcc using rendering-assisted distillation technology

Original title: Radocc: Learning Cross-Modality Occupancy Knowledge through Rendering Assisted Distillation

Paper link: https://arxiv.org/pdf/2312.11829.pdf

Author affiliation: FNii , CUHK-Shenzhen SSE, CUHK-Shenzhen Huawei Noah's Ark Laboratory

Conference: AAAI 2024

Paper idea:

3D occupancy prediction is an emerging task that aims to estimate the occupancy status and semantics of 3D scenes using multi-view images. However, image-based scene perception encounters significant challenges in achieving accurate predictions due to the lack of geometric priors. This paper addresses this problem by exploring cross-modal knowledge distillation in this task, i.e., we utilize a more powerful multi-modal model to guide the visual model during the training process. In practice, this paper observes that direct application of feature or logits alignment, proposed and widely used in bird's-eye view (BEV) perception, does not yield satisfactory results. To overcome this problem, this paper introduces RadOcc, a rendering-assisted distillation paradigm for 3D occupancy prediction. By employing differentiable volume rendering, we generate depth and semantic maps in perspective and propose two novel consistency criteria between the rendered output of teacher and student models. Specifically, the depth consistency loss aligns the termination distributions of rendering rays, while the semantic consistency loss mimics the intra-segment similarity guided by the visual base model (VLM). Experimental results on the nuScenes dataset demonstrate the effectiveness of the method proposed in this article in improving various 3D occupancy prediction methods. For example, the method proposed in this article improves the baseline of this article by 2.2% in the mIoU metric and reaches 2.2% in the Occ3D benchmark. 50%.

Main contributions:

This paper introduces a rendering-assisted distillation paradigm called RadOcc for 3D occupancy prediction. This is the first paper exploring cross-modal knowledge distillation in 3D-OP, providing valuable insights into the application of existing BEV distillation techniques in this task.

The authors propose two novel distillation constraints, namely rendering depth and semantic consistency (RDC and RSC). These constraints effectively enhance the knowledge transfer process by aligning light distribution and correlation matrices guided by the vision base model. The key to this approach is to use depth and semantic information to guide the rendering process, thereby improving the quality and accuracy of the rendering results. By combining these two constraints, the researchers achieved significant improvements, providing new solutions for knowledge transfer in vision tasks.

Equipped with the proposed method, RadOcc shows state-of-the-art dense and sparse occupancy prediction performance on Occ3D and nuScenes benchmarks. In addition, experiments have proven that the distillation method proposed in this article can effectively improve the performance of multiple baseline models.

Network design:

This paper is the first to study cross-modal knowledge distillation for the 3D occupancy prediction task. Based on the method of knowledge transfer using BEV or logits consistency in the BEV sensing field, this paper extends these distillation techniques to the 3D occupancy prediction task, aiming to align voxel features and voxel logits, as shown in Figure 1(a) . However, preliminary experiments show that these alignment techniques face significant challenges in 3D-OP tasks, especially the former method that introduces negative transfer. This challenge may stem from the fundamental difference between 3D object detection and occupancy prediction, which as a more fine-grained perception task requires capturing geometric details as well as background objects.

To address the above challenges, this paper proposes RadOcc, a novel method for cross-modal knowledge distillation using differentiable volume rendering. The core idea of ​​RadOcc is to align the rendering results generated by the teacher model and the student model, as shown in Figure 1(b). Specifically, this article uses the intrinsic and extrinsic parameters of the camera to perform volume rendering of voxel features (Mildenhall et al. 2021), which enables this article to obtain corresponding depth maps and semantic maps from different viewpoints. To achieve better alignment between rendered outputs, this paper introduces novel Rendering Depth Consistency (RDC) and Rendering Semantic Consistency (RSC) losses. On the one hand, RDC loss enforces the consistency of ray distribution, which enables the student model to capture the underlying structure of the data. On the other hand, the RSC loss takes advantage of the visual base model (Kirillov et al. 2023) and utilizes pre-extracted segments for affinity distillation. This standard allows models to learn and compare semantic representations of different image regions, thereby enhancing their ability to capture fine-grained details. By combining the above constraints, the method proposed in this paper effectively leverages cross-modal knowledge distillation, thereby improving performance and better optimizing the student model. This paper demonstrates the effectiveness of our approach on dense and sparse occupancy prediction, achieving state-of-the-art results on both tasks.

Figure 1: Render-assisted distillation. (a) Existing methods align features or logits. (b) The RadOcc method proposed in this paper simultaneously constrains the rendered depth map and semantics. Figure 2: Overall framework of RadOcc. It adopts a teacher-student architecture, where the teacher network is a multi-modal model and the student network only accepts camera input. The predictions of both networks will be used to generate rendering depth and semantics through differentiable volume rendering. Newly proposed rendering depth and semantic consistency losses are adopted between rendering results.

Figure 3: Rendering depth analysis. Although the teacher (T) and student (S) have similar rendering depths, especially for foreground objects, their light termination distributions show large differences.

Figure 4: Generation of affinity matrix. This article first uses the Vision Foundation Model (VFM), namely SAM, to extract segments into the original image. Afterwards, this article performs segment aggregation on the semantic features rendered in each segment to obtain the affinity matrix.

Experimental results:

##Summary:

This paper proposes RadOcc, a new cross-modal approach for 3D occupancy prediction Knowledge distillation paradigm. It utilizes a multimodal teacher model to provide geometric and semantic guidance to the visual student model through differentiable volume rendering. Furthermore, this paper proposes two new consistency criteria, depth consistency loss and semantic consistency loss, to align the ray distribution and affinity matrix between teacher and student models. Extensive experiments on Occ3D and nuScenes datasets show that RadOcc can significantly improve the performance of various 3D occupancy prediction methods. Our method achieves state-of-the-art results on the Occ3D challenge benchmark and significantly outperforms existing published methods. We believe that our work opens up new possibilities for cross-modal learning in scene understanding.

The above is the detailed content of Learning cross-modal occupancy knowledge: RadOcc using rendering-assisted distillation technology. For more information, please follow other related articles on the PHP Chinese website!