DeepUSPS: Deep Robust Unsupervised Saliency Prediction via Self-supervision
Authors: Tam Nguyen, Maximilian Dax, Chaithanya Kumar Mummadi, Nhung Ngo, Thi Hoai Phuong Nguyen, Zhongyu Lou, Thomas Brox
NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 4 Experiments We first compare our proposed pipeline to existing benchmarks by following the configuration of Zhang et al. (2018). Further, we show in detailed oracle and ablation studies how each component of the pipeline is crucial for the overall competitive performance. |
| Researcher Affiliation | Collaboration | Equal contribution, [fixed-term.Maximilian.Dax, Ductam.Nguyen]@de.bosch.com Computer Vision Group, University of Freiburg, Germany Bosch Research, Bosch Gmb H, Germany Bosch Center for AI, Bosch Gmb H, Germany Karlsruhe Institute of Technology, Germany |
| Pseudocode | No | The paper describes its methods through text and diagrams (e.g., Fig. 3, Fig. 4) but does not include structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code is available at https://tinyurl.com/wtlhgo3 . |
| Open Datasets | Yes | Our method is evaluated on traditional object saliency prediction benchmarks (Borji et al., 2015). Following Zhang et al. (2018), we extract handcrafted maps from MSRA-B (Liu et al., 2010): 2500 and 500 training and validation images respectively. The remaining test set contains in total 2000 images. Further tests are performed on the ECCSD-dataset (Yan et al., 2013) (1000 images), DUT (Yang et al., 2013) (5168 images), SED2 (Alpert et al., 2011)(100 images). |
| Dataset Splits | Yes | Following Zhang et al. (2018), we extract handcrafted maps from MSRA-B (Liu et al., 2010): 2500 and 500 training and validation images respectively. The remaining test set contains in total 2000 images. |
| Hardware Specification | Yes | Our proposed pipeline needs about 30 hours of computation time on four Geforce Titan X for training. |
| Software Dependencies | No | The paper mentions software components and tools like ADAM, DRN-network, ResNet101, and fully-connected CRF, but does not specify their version numbers or other crucial software dependencies required for replication. |
| Experiment Setup | Yes | Our pseudo generation networks are trained for a fixed number of 25 epochs for each handcrafted method and saliency detection network is trained for 200 epochs in the final stage.We use ADAM (Kingma & Ba, 2014) with a momentum of 0.9, batch size 20, a learning rate of 1e-6 in the first step when trained on the handcrafted methods. The learning rate is doubled every time in later self-supervision iteration. |