We introduce the task of 3D visual grounding in large-scale dynamic scenes based on natural linguistic descriptions and online captured multi-modal visual data, including 2D images and 3D LiDAR point clouds. We present a novel method, dubbed WildRefer, for this task by fully utilizing the rich appearance information in images, the position and geometric clues in point cloud as well as the semantic knowledge of language descriptions. Besides, we propose two novel datasets, i.e., STRefer and LifeRefer, which focus on large-scale human-centric daily-life scenarios accompanied with abundant 3D object and natural language annotations. Our datasets are significant for the research of 3D visual grounding in the wild and has huge potential to boost the development of autonomous driving and service robots. Extensive experiments and ablation studies demonstrate that our method achieves state-of-the-art performance on the proposed benchmarks.
Pipeline of WildRefer. The inputs are multi-frame synchronized points clouds and images as well as a natural language description. After feature extraction, we obtain two types of visual feature and a text feature. Through two dynamic visual encoders, we extract dynamic-enhanced image and point features. Then, a triple-modal feature interaction module is followed to fuse valuable information from different modalities. Finally, through a DETR-like decoder, we decode the location and size of the target object. SA and CA denote self-attention and cross-attention, respectively.
STRefer
LifeRefer
STRefer & LifeRefer |── group_id # strefer or liferefer |── scene_id # unique scene id that can match STCrowd & HucenLife |── object_id # unique object id (For LifeRefer, it is kept same with HucenLife. | For STRefer, it consist of scene id, frame id and object id.) |── point_cloud | |── point_cloud_name # the frame name of point cloud for the scene | |── bbox # bounding box of the object | |── category # category of the object |── language | |── description # language description of the object | |── token # token of the description | |── ann_id # annotation id of the object |── image | |── image_name # the frame name of image for the scene |── calibration | |── ex_matrix # external matrix of calibration | |── in_matrix # internal matrix of calibration
We strongly recommend use our preproceed data of STCrowd and HucenLife.
Comparison results on STRefer and LifeRefer. {*} denotes the one-stage version without pretrained 3D decoder.
Method | Publication | Type | STRefer Acc@0.25 |
STRefer Acc@0.5 |
STRefer mIOU |
LifeRefer Acc@0.25 |
LifeRefer Acc@0.5 |
LifeRefer mIOU |
Time cost (ms) |
---|---|---|---|---|---|---|---|---|---|
ScanRefer | ECCV-2020 | Two-Stage | 32.93 | 30.39 | 25.21 | 22.76 | 14.89 | 12.61 | 156 |
ReferIt3D | ECCV-2020 | Two-Stage | 34.05 | 31.61 | 26.05 | 26.18 | 17.06 | 14.38 | 194 |
3DVG-Transformer | ICCV-2021 | Two-Stage | 40.53 | 37.71 | 30.63 | 25.71 | 15.94 | 13.99 | 160 |
MVT | CVPR-2022 | Two-Stage | 45.12 | 42.40 | 35.03 | 21.65 | 13.17 | 11.71 | 242 |
3DJCG | CVPR-2022 | Two-Stage | 50.47 | 47.47 | 38.85 | 27.82 | 16.87 | 15.40 | 161 |
BUTD-DETR | ECCV-2022 | Two-Stage | 57.60 | 47.47 | 35.22 | 30.81 | 11.66 | 14.80 | 252 |
EDA | CVPR-2023 | Two-Stage | 55.91 | 47.28 | 34.32 | 31.44 | 11.18 | 15.00 | 291 |
3D-SPS | CVPR-2022 | One-Stage | 44.47 | 42.40 | 30.43 | 28.01 | 18.20 | 15.78 | 130 |
BUTD-DETR* | ECCV-2022 | One-Stage | 56.66 | 45.12 | 33.52 | 32.46 | 12.82 | 15.96 | 138 |
EDA* | CVPR-2023 | One-Stage | 57.41 | 45.59 | 34.03 | 29.32 | 12.25 | 14.41 | 154 |
Ours | – | One-Stage | 62.01 | 54.97 | 38.77 | 38.89 | 18.42 | 19.47 | 151 |