Chinese Journal of Liquid Crystals and Displays, Volume. 39, Issue 9, 1243(2024)
Prototype distribution correction for few-shot point cloud classification
Fig. 3. Visualization of feature distributions before and after prototype correction
Fig. 4. Impact of the number of generated features on classification accuracy
Table 1. Few-shot classification results on ModelNet40 and ModelNet40-C
View tableView in ArticleTable 1. Few-shot classification results on ModelNet40 and ModelNet40-C
方法 网络 ModelNet40 ModelNet40-C 5way-1shot 5way-5shot 5way-1shot 5way-5shot MetaOptNet［ 22 ］DGCNN 75.77±0.83 86.44±0.62 73.34±0.85 85.15±0.63 RelationNet［ 25 ］DGCNN 77.46±0.80 85.11±0.61 — — ProtoNet［ 7 ］DGCNN 79.46±0.76 88.65±0.54 77.69±0.77 86.81±0.58 CIA［ 5 ］DGCNN 83.46±0.83 89.15±0.50 80.64±0.86 88.23±0.54 Cross-Modality［ 8 ］DGCNN* 83.89 ±0.75 90.64±0.52 81.87 ±0.78 89.51±0.54 Proto-Correction DGCNN_O 83.66±0.78 90.70 ±0.51 80.79±0.80 90.32 ±0.51 Top-k DGCNN_O 83.66±0.78 90.59±0.50 80.73±0.80 90.29±0.51
Table 2. Few-shot classification results on ScanObjectNN and ScanObjectNN-PB
View tableView in ArticleTable 2. Few-shot classification results on ScanObjectNN and ScanObjectNN-PB
方法 网络 ScanObjectNN ScanObjectNN-PB 5way-1shot 5way-5shot 5way-1shot 5way-5shot MetaOptNet［ 22 ］DGCNN 61.12±0.66 67.73±0.45 57.15±0.63 65.56±0.50 RelationNet［ 25 ］DGCNN 54.23±0.63 66.72±0.50 — — ProtoNet［ 7 ］DGCNN 60.46±0.67 70.20±0.52 59.29±0.65 67.68±0.47 CIA［ 5 ］DGCNN 62.17±0.68 71.31±0.45 57.02±0.68 67.37±0.49 Cross-Modality［ 8 ］DGCNN* 64.69±0.64 74.60±0.43 60.25±0.63 71.00±0.47 Proto-Correction DGCNN_O 66.38 ±0.69 76.53 ±0.46 61.90 ±0.68 71.14±0.48 Top-k DGCNN_O 66.34±0.70 76.52±0.47 61.85±0.69 71.15 ±0.48
Table 3. Effects of orthogonal constraints for network parameters
View tableView in ArticleTable 3. Effects of orthogonal constraints for network parameters
方法 ModelNet40-1shot ScanObjectNN-1shot DGCNN DGCNN_O DGCNN DGCNN_O MetaOptNet［ 22 ］75.77±0.83 78.85±0.75（↑3.08） 61.12±0.66 61.72±0.58（↑0.60） RelationNet［ 25 ］77.46±0.80 — 54.23±0.63 — ProtoNet［ 7 ］79.46±0.76 80.98±0.72（↑1.52） 60.46±0.67 63.73±0.59（↑3.27） CIA［ 5 ］83.46±0.83 84.31±0.79（↑0.85） 62.17±0.68 63.04±0.59（↑0.87） Cross-Modality［ 8 ］83.89±0.75 84.01±0.73（↑0.12） 64.69±0.64 65.16±0.65（↑0.47） Proto-Correction 80.70±0.86 83.66±0.78（↑2.96） 62.01±0.62 66.38±0.69（↑4.37） Top-k 82.01±0.77 83.66±0.78（↑1.65） 63.05±0.62 66.34±0.70（↑3.29）
Table 4. Effects of prototype correction method
View tableView in ArticleTable 4. Effects of prototype correction method
Proto-Correction Top-k ModelNet40 ScanObjectNN 5way-1shot 5way-5shot 5way-1shot 5way-5shot × × 80.98±0.72 90.37±0.50 63.73±0.59 75.88±0.45 √ × 83.66±0.78 90.70 ±0.51 66.38 ±0.69 76.53 ±0.46 √ √ 83.66 ±0.78 90.59±0.50 66.34±0.70 76.52±0.47
Get Citation
Copy Citation Text
Yuanzhi FENG, Yu XIA, Jielong GUO, Dongheng SHAO, Jianfeng ZHANG, Xian WEI. Prototype distribution correction for few-shot point cloud classification[J]. Chinese Journal of Liquid Crystals and Displays, 2024, 39(9): 1243
Category:
Received: Nov. 22, 2023
Accepted: --
Published Online: Nov. 13, 2024
The Author Email: Xian WEI (xian.wei@fjirsm.ac.cn)
