論文/代碼速遞2022.10.13！
最新成果demo展示：
主頁：https://sgvr.kaist.ac.kr/publication/flow-supervisor/ 代碼：https://github.com/iwbn/flow-supervisor
光流CNN的訓(xùn)練管道由合成數(shù)據(jù)集的預(yù)訓(xùn)練階段和目標(biāo)數(shù)據(jù)集的微調(diào)階段組成。然而，從目標(biāo)視頻中獲取ground truth 流需要付出巨大的努力。本文提出了一種實用的微調(diào)方法，以使預(yù)處理模型適應(yīng)沒有g(shù)round truth 流的目標(biāo)數(shù)據(jù)集，這種方法尚未得到廣泛的探索。具體來說，我們提出了一個用于自監(jiān)督的流監(jiān)督，它由參數(shù)分離和學(xué)生輸出連接組成。這種設(shè)計的目的是穩(wěn)定收斂，并比在微調(diào)任務(wù)中不穩(wěn)定的傳統(tǒng)自監(jiān)督方法具有更好的精度。實驗結(jié)果表明，與不同的自監(jiān)督方法相比，該方法對于半監(jiān)督學(xué)習(xí)是有效的。此外，通過利用額外的未標(biāo)記數(shù)據(jù)集，我們在Sintel和KITTI基準(zhǔn)上對最先進(jìn)的光流模型進(jìn)行了有意義的改進(jìn)
最新論文整理
ECCV2022
Updated on : 13 Oct 2022
total number : 1
DeepMend: Learning Occupancy Functions to Represent Shape for Repair
論文/Paper: http://arxiv.org/pdf/2210.05728
代碼/Code: https://github.com/terascale-all-sensing-research-studio/deepmend
CVPR2022
NeurIPS
Updated on : 13 Oct 2022
total number : 11
AniFaceGAN: Animatable 3D-Aware Face Image Generation for Video Avatars
論文/Paper: http://arxiv.org/pdf/2210.06465
代碼/Code: None
Learning Multi-resolution Functional Maps with Spectral Attention for Robust Shape Matching
論文/Paper: http://arxiv.org/pdf/2210.06373
代碼/Code: https://github.com/craigleili/AttentiveFMaps
Latency-aware Spatial-wise Dynamic Networks
論文/Paper: http://arxiv.org/pdf/2210.06223
代碼/Code: https://github.com/leaplabthu/lasnet
Multi-Granularity Cross-modal Alignment for Generalized Medical Visual Representation Learning
論文/Paper: http://arxiv.org/pdf/2210.06044
代碼/Code: None
Long-Form Video-Language Pre-Training with Multimodal Temporal Contrastive Learning
論文/Paper: http://arxiv.org/pdf/2210.06031
代碼/Code: https://github.com/microsoft/XPretrain.
Boosting the Transferability of Adversarial Attacks with Reverse Adversarial Perturbation
論文/Paper: http://arxiv.org/pdf/2210.05968
代碼/Code: https://github.com/sclbd/transfer_attack_rap
Decomposed Knowledge Distillation for Class-Incremental Semantic Segmentation
論文/Paper: http://arxiv.org/pdf/2210.05941
代碼/Code: None
A Lower Bound of Hash Codes' Performance
論文/Paper: http://arxiv.org/pdf/2210.05899
代碼/Code: https://github.com/vl-group/lbhash
SegViT: Semantic Segmentation with Plain Vision Transformers
論文/Paper: http://arxiv.org/pdf/2210.05844
代碼/Code: None
Trap and Replace: Defending Backdoor Attacks by Trapping Them into an Easy-to-Replace Subnetwork
論文/Paper: http://arxiv.org/pdf/2210.06428
代碼/Code: https://github.com/VITA-Group/Trap-and-Replace-Backdoor-Defense
Towards Theoretically Inspired Neural Initialization Optimization
論文/Paper: http://arxiv.org/pdf/2210.05956
代碼/Code: https://github.com/HarborYuan/GradCosine