Learning Equivariant Segmentation with Instance-Unique Querying

Authors: Wenguan Wang, James Liang, Dongfang Liu

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental On top of four famous, query-based models (i.e., Cond Inst, SOLOv2, SOTR, and Mask2Former), our training algorithm provides significant performance gains (e.g., +1.6 3.2 AP) on COCO dataset. In addition, our algorithm promotes the performance of SOLOv2 by 2.7 AP, on LVISv1 dataset.
Researcher Affiliation Academia Wenguan Wang Re LER, AAII, University of Technology Sydney James Liang Rochester Institute of Technology Dongfang Liu Rochester Institute of Technology
Pseudocode Yes The pseudo-code is also included in the supplemental material.
Open Source Code Yes https://github.com/James Liang819/Instance_Unique_Querying ... our full implementations are made publicly available at https://github.com/James Liang819/Instance_Unique_Querying.
Open Datasets Yes We conduct our main experiments on the gold-standard benchmark dataset, i.e., COCO[47], in this field. ... we perform additional experiments on LVISv1 [48] on top of SOLOv2 [39]. ... The datasets used in this work are online publicly available and well cited in the paper. MS COCO dataset is public and licensed under a Creative Commons Attribution 4.0 License.
Dataset Splits Yes We use train2017 split (115k images) for training and val2017 (5k images) for validation used in our ablation study. ... it contains a total of 100k images of a train set under a significant long-tailed distribution, and relatively balanced val (20k images) and test sets (20k images).
Hardware Specification Yes For all our experiments, the training and testing are conducted on eight NVIDIA Tesla A100 GPUs with a 80GB memory per-card.
Software Dependencies No The paper mentions that its algorithm is 'implemented in Py Torch' and based on 'Adelai Det [76] and MMDetection [77]', and mentions specific licenses for used codes. However, it does not provide specific version numbers for PyTorch, Adelai Det, MMDetection, or any other core software libraries/solvers.
Experiment Setup Yes We train models using SGD with initial learning rate of 0.01 for Res Net [45] backboned models and Adam with initial learning rate of 1e 5 for Swin [46] backboned models. The learning rate is scheduled following the polynomial annealing policy with batch size 16. ... Res Net backboned models are trained for 12 epochs, while Swin backboned models are trained for 50 epochs. ... We implement Linter_mask (cf. Eq. 5) as the focal loss [66], whose hyper-parameters are set to α = 0.1 and γ = 2.5. ... To balance the impacts of our two new training targets, i.e., Linter_mask and Lequi, we multiply Lequi by a coefficient λ, which is empirically set as 3.