simCrossTrans: A Simple Cross-Modality Transfer Learning for Object Detection with ConvNets or Vision Transformers

Transfer learning is widely used in computer vision (CV), natural languageprocessing (NLP) and achieves great success. Most transfer learning systems arebased on the same modality (e.g. RGB image in CV and text in NLP). However, thecross-modality transfer learning (CMTL) systems are scarce. In this work, westudy CMTL from 2D to 3D sensor to explore the upper bound performance of 3Dsensor only systems, which play critical roles in robotic navigation andperform well in low light scenarios. While most CMTL pipelines from 2D to 3Dvision are complicated and based on Convolutional Neural Networks (ConvNets),ours is easy to implement, expand and based on both ConvNets and Visiontransformers(ViTs): 1) By converting point clouds to pseudo-images, we can usean almost identical network from pre-trained models based on 2D images. Thismakes our system easy to implement and expand. 2) Recently ViTs have beenshowing good performance and robustness to occlusions, one of the key reasonsfor poor performance of 3D vision systems. We explored both ViT and ConvNetwith similar model sizes to investigate the performance difference. We name ourapproach simCrossTrans: simple cross-modality transfer learning with ConvNetsor ViTs. Experiments on SUN RGB-D dataset show: with simCrossTrans we achieve$13.2\%$ and $16.1\%$ absolute performance gain based on ConvNets and ViTsseparately. We also observed the ViTs based performs $9.7\%$ better than theConvNets one, showing the power of simCrossTrans with ViT. simCrossTrans withViTs surpasses the previous state-of-the-art (SOTA) by a large margin of$+15.4\%$ mAP50. Compared with the previous 2D detection SOTA based RGB images,our depth image only system only has a $1\%$ gap. The code, training/inferencelogs and models are publicly available athttps://github.com/liketheflower/simCrossTrans