基于知识图谱引导的三维视觉问答

毛爱华,陈思宇

3D visual question answering guided by knowledge graph

Aihua MAO,Siyu CHEN

表 1 不同方法在ScanQA带有对象的测试集上的评估指标结果对比

Tab.1 Comparison evaluation metrics for different methods on ScanQA test set with objects

方法 EM@1/% EM@10/% BLEU-1/% BLEU-4/% ROUGE/% METEOR/% CIDEr RandomImage+MCAN[1] 22.31 53.11 26.66 14.26 31.27 12.13 60.37 VoteNet+MCAN[1] 19.71 50.76 29.46 6.08 30.97 12.07 58.23 scanRefer+MCAN(e2e) [1] 20.56 52.35 27.85 7.46 30.68 11.97 57.36 scanQA[1] 23.45 56.51 31.56 12.04 34.34 13.55 67.29 3D-VisTA(scratch)[11] 25.20 55.20 — 10.50 35.50 13.80 68.60 Multi-CLIP[2] 22.76 — 31.08 13.31 33.84 13.28 65.81 本研究 24.47 56.18 33.85 13.59 36.26 14.42 71.33 jin-Context-aware*[4] 24.58 55.97 33.15 11.23 35.97 14.16 70.18 Multi-CLIP(pre-trained)*[2] 24.02 — 32.63 12.65 35.46 13.97 68.70 3D-VisTA*[11] 27.00 57.90 — 16.00 38.60 15.20 76.60 CLIP*[3] 23.92 — 32.72 14.64 35.15 13.94 69.53