Make One-Shot Video Object Segmentation Efficient Again
Authors: Tim Meinhardt, Laura Leal-Taixé
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the applicability of e-OSVOS on three semi-supervised VOS benchmarks, namely, DAVIS 2016 [27], DAVIS 2017 [28] and You Tube-VOS [41]. |
| Researcher Affiliation | Academia | Tim Meinhardt Technical University of Munich tim.meinhardt@tum.de Laura Leal-Taixé Technical University of Munich leal.taixe@tum.de |
| Pseudocode | Yes | In Algorithm 1 of the supplementary, we illustrate the full e-OSVOS training pipeline for a given VOS taskset Ttrain. |
| Open Source Code | Yes | Code is available at https://github.com/dvl-tum/e-osvos. |
| Open Datasets | Yes | We demonstrate the applicability of e-OSVOS on three semi-supervised VOS benchmarks, namely, DAVIS 2016 [27], DAVIS 2017 [28] and You Tube-VOS [41]. ... Mask R-CNN model with Res Net50 [12] and FPN [17] pre-trained on the COCO [18] segmentation dataset. |
| Dataset Splits | Yes | DAVIS 2016 The DAVIS 2016 [27] benchmark consists of a training and validation set with 30 and 20 single object video sequences, respectively. ... DAVIS 2017 The DAVIS 2017 [28] benchmark extends DAVIS-16 with 100 additional sequences including dedicated test-dev and test sets. The validation, test-dev and test sets each consist of 30 sequences. ... You Tube-VOS Our largest benchmark, You Tube-VOS [41], consists of 4453 video sequences including dedicated test and validation sets with 508 and 474 sequences, respectively. |
| Hardware Specification | Yes | each distributed to a Quadro RTX 6000 GPU for a total of 4 days. |
| Software Dependencies | No | The paper mentions Mask R-CNN, ResNet50, FPN, and RAdam, but does not provide specific version numbers for these or other software dependencies. |
| Experiment Setup | Yes | To limit the computational effort, we ignore second order derivatives and fine-tune for T = 5 BPTT iterations. ... The online adaptation (On A) is applied every IOn A = 5 steps for TOn A = 10 iterations. |