readQueueP1.mdpp

### First pass papers
This file contains an overview of the papers I recently red (only superficially in one pass). Sorted by time of reading in reverse order. (✓✓) only given if the observed parts of the paper are extremely well written.


#### DATE SUBJECT

##### Notes

##### Q



#### `2018-10-02` SUBJECT

##### Notes

##### Q

```latex
@inproceedings{bhattacharyya2018long,
  title={Long-term on-board prediction of people in traffic scenes under uncertainty},
  author={Bhattacharyya, Apratim and Fritz, Mario and Schiele, Bernt},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  pages={4194--4202},
  year={2018}
}
```


#### `2018-10-02` SUBJECT (revisit)

##### Notes

##### Q

```latex
@article{yagi2017future,
  title={Future Person Localization in First-Person Videos},
  author={Yagi, Takuma and Mangalam, Karttikeya and Yonetani, Ryo and Sato, Yoichi},
  journal={arXiv preprint arXiv:1711.11217},
  year={2017}
}
```


#### `2018-10-02` Siamese Networks (possibly revisit)

##### Q
- category: proposes Siamese NN for facial verification
- Siamese Networks
  - use cases: 
    - recognition + verification
    - number training / class can be very low
    - classes do not need to be known during training
  - not useful alternatives
    1. Distance based methods (compute similarity metric between pattern and lib of stored prototypes)
    2. Generative models in reduced dimension space
  - idea: 
    - learn f : INPUT -> TARGET sth \| f(input) \|_1 maintains semantic 
      distance in input space
    - mapping f := convolutional NN

```latex
@inproceedings{chopra2005learning,
  title={Learning a similarity metric discriminatively, with application to face verification},
  author={Chopra, Sumit and Hadsell, Raia and LeCun, Yann},
  booktitle={Computer Vision and Pattern Recognition, 2005. CVPR 2005. IEEE Computer Society Conference on},
  volume={1},
  pages={539--546},
  year={2005},
  organization={IEEE}
}
```


#### `2018-10-01` Siamese Networks 
##### Notes
- `one shot learning` means to predict on the basis of only one single example of each new class
  - normally; for one shot learning, domain specific features are designed
    - works fine on similar inputs
    - does however not offer robust solution
- Siamese networks (see DeepFace paper)
  - := NN-architecture that contains two or more identical subnetworks model architecture (involves distance-based loss)
    - learns similarity of two pairs of inputs
    - training thus involves pairwise learning -> quadratic pairs to learn
  - advantages
    - more robust to class imbalance
    - good for one shot learning
    - learn similarity
  - Feature extraction
  - employ unique structure to naturally rank similarity between inputs

##### Q
- contribution: 
  1. Method for learning in `siamese neural networks`
  
  
```latex
@inproceedings{koch2015siamese,
  title={Siamese neural networks for one-shot image recognition},
  author={Koch, Gregory and Zemel, Richard and Salakhutdinov, Ruslan},
  booktitle={ICML Deep Learning Workshop},
  volume={2},
  year={2015}
}
```

#### `2018-10-01` Gaze Anticipation (revisit)
##### Notes
- point of gaze := point which human is fixating
- method (called Deep Future Gaze := `DFG`): 
  - GAN
    - GENERATOR: 
        1. (PREDICT FRAMES) Two-stream spatial temporal convolutional net 
           (3d-CNN); one for background; one for foreground 
        2. Second 3d-CNN employed for gaze anticipations on top 
    - DISCRIMINATOR
    - This GAN is used for predicting the future frames
- [Experimental results and code](https://github.com/Mengmi/deepfuturegaze_gan    )

##### Q
- contribution 
  1. Aforementioned method outperforms state-of-the-art methods
    - prediction of the point of gaze in the future (order of a few seconds)
  2. Introduces new egocentric dataset 
- reproducibility: ✓✓
- clarity: ✓
- correctness: ✓ even though I am extremely surprised that the prediction of the future frames in the next few seconds can be done reliably (gaze anticipation is stacked on top of that)

```latex
@inproceedings{zhang2017deep,
  title={Deep Future Gaze: Gaze Anticipation on Egocentric Videos Using Adversarial Networks.},
  author={Zhang, Mengmi and Ma, Keng Teck and Lim, Joo-Hwee and Zhao, Qi and Feng, Jiashi},
  booktitle={CVPR},
  pages={3539--3548},
  year={2017}
}
```

#### `2018-10-01`  Video summarization (revisit)
##### Notes
- video summarization := "compress" video by extracting important parts
  - viewpoint := yields `important` aspects of the image / video
  - applying the video summarization we receive similarity metrics wrt the
    viewpoint.
  - not one optimal viewpoint

##### Q
- contribution meta:
  - viewpoint specified as `similarity` (in a semantic sense) which, once
    again is depending on the aspect that is considered important.
  - focuses on video-level similarities (supervized) 
- core contribution:
  1. Proposes a novel video summarization method from multiple groups; inspired
     by Fisher's discriminant 
  2. Introduces a dataset for that purpose
  3. Developed optimization algorithm t
- usefulness:
  - can be used for feature extraction
- clarity: ✓✓
- correctness: ✓

```latex
@inproceedings{kanehira2018aware,
  title={aware Video Summarization},
  author={Kanehira, Atsushi and Van Gool, Luc and Ushiku, Yoshitaka and Harada, Tatsuya},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  pages={7435--7444},
  year={2018}
}
```