Combinatorial Optimization for Panoptic Segmentation: A Fully Differentiable Approach

Authors: Ahmed Abbas, Paul Swoboda

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental evaluation shows improvement by backpropagating through the optimization problem w.r.t. comparable approaches on Cityscapes and COCO datasets.
Researcher Affiliation Academia Ahmed Abbas Paul Swoboda MPI for Informatics Saarland Informatics Campus
Pseudocode Yes Algorithm 1: BACKWARD PASS
Open Source Code Yes Our code is available at https://github.com/aabbas90/COPS.
Open Datasets Yes We train and evaluate COPS on the Cityscapes [19] and COCO [47] panoptic segmentation datasets.
Dataset Splits Yes Cityscapes: Contains traffic related images of resolution 1024 2048 where training, validation and testing splits have 2975, 500, and 1525 images for training, validation, and testing, respectively.
Hardware Specification Yes All baselines are trained on NVIDIA Quadro RTX 8000 GPUs with 48GB memory each. For fully differentiable training we use one Tesla P40 with 24GB memory and a 32 core CPU to solve all AMWC problems in the batch in parallel.
Software Dependencies No The paper states, 'We closely follow the implementation of Panoptic-Deep Lab in [70] (based on Pytorch [54])', but it does not provide specific version numbers for Pytorch or any other software dependency.
Experiment Setup Yes During training we use random scale augmentation and crop to 512 1024 resolution as done in Panoptic-Deep Lab. The values of small segment rejection thresholds (used during both training and inference) are 200, 2048 for thing and stuff class resp. We train Cityscapes on one GPU with batch-size 12 for 250k iterations, with initial learning rate 0.001. COCO is trained on four GPUs with a total batch-size of 48 for 240k iterations. We train with batch size of 24 until training loss convergences which amounts to 3000 iterations for Cityscapes and 10000 iterations for COCO. We use N = 5 in our experiments.