Skip to content

Latest commit

 

History

History

models

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
analysis
analysis
 
 
generalizability
generalizability
 
 
scratch
scratch
 
 
README.md
README.md
 
 

README.md

Models

Our paper have multiple analysis for different settings. Here we release the model description, and trained models for each of them. Our analysis are based on VGG-16 architecture.

Section 1: Analysis

In this section, we consider a single-scale 224x224 input image and used two different tasks (a). semantic segmentation on PASCAL VOC-2012; and (b). surface normal estimation on NYU-depth dataset.

Here we show how to train linear and non-linear models effectively using skip connections.

Section 2: Scratch

We also trained our model from scratch for semantic segmentation and surface normal estimation. Since surface normal estimation (or as we call Geometry in our paper) does not require any human labels, it can also be considered as a representation learnt in a self-supervised manner. For self-supervised representation learning, we evaluated our approach for semantic segmentation on PASCAL VOC-2012 and object detection on PASCAL VOC-2007.

We found batch-normalization as an important tool along with our simple technique of sampling to train linear models, and training from scratch.

Section 3: Generalizability

Finally, we release the models for different tasks trained for different datasets:

  1. PASCAL Context: We trained models for PASCAL Context dataset for 59 and 39 classes. The evaluation is done on the validation set.

  2. PASCAL VOC-2012: We trained models for PASCAL VOC 2012 data for 21 classes (20 categories + background). We augmented our dataset using the annotations provided by Hariharan et al. The model was evaluated on the test set of PASCAL VOC-2012 using the evaluation server.

  3. NYUv2 Surface Normal: We trained models for surface normal estimation using 795 unique trainval sequences (containing kinect data for each RGB frame). For details about dataset and how to download it, please refer to MarrRevisited paper.

  4. BSDS Edge Detection: We trained models for edge detection on BSDS-500 dataset.

Please consider citing the above papers in case you use any of the above dataset or models trained using them.

Contact

Please contact Aayush Bansal if some model is missing or you want to know about some aditional analysis.