Indoor scene recognition task entry Raiders! First experience sharing, Malena.

Indoor scene recognition task entry Raiders! First experience sharing, Malena.

First, the problem of the game 

Indoor scene recognition in the field of computer vision is a challenging problem, most suitable for outdoor scene classification model performance in indoor areas than the poor, its difficulty is that although some indoor scenes can be better characterized by global space described, but other scenarios need specific features of the object included in the scene can be better described.

Therefore, how to extract local features and global features image to accurately predict the difficulty of the game is a problem.

 

• In this data set includes a total of 67 indoor scene category, a total of 15,620 images, each category has at least 100 images, where the training set, validation set and test set into a ratio of 6: 2: 2. 
• Data set pictures in larger quantities, while observing 100 image size generally concentrated in a similar size of 300 * 400, but there are also individual 1000 * 1000 and more moderate consider when choosing the size of 512 * 512 more appropriate 

 

Second, the core ideas 

1. Because indoor scene classification data sets, we can go to find training methods related data sets, you can find similar place205, place365 and other similar interior scenes classified data set for pre-training weights model training heavy, you can load pre-training weight training, results the relative would be better, but later found, in fact, simply just too heavy can get almost the same effect with the right imagenet pre-trained, most of the credit is due to the adjustment parameters.
2. Parameter adjustment is a very important part of the data preprocessing attempted vertical and horizontal flipping, effect is reduced, brightness adjustment attempt, the fuzzy processing operations, the effect is decreased, so there is no enhancement data simply.
3. During training, the use of cosine annealing way to reduce the learning rate, in the middle also need to be adjusted very many times before they can find a relatively appropriate learning rate.
4. The final step is the replacement model, the best beginning to try a big difference model to determine a base model making changes. Finally, the operation test of time, you can add data enhancement, may be fusion model, a variety of ways you can go try, with mixed success. 

 

Third, the game experience summary 

• Data set is large, the training time is relatively long in advance to choose the right model for training is the most important;
• After selecting the appropriate model similar to the model in the middle of a finer selection, some data sets really eat architectural model, so be careful to select the model, the wrong model limits the performance improvements, wrong step back is a waste of time;
• If the ability to test, then you can add enhanced at the time of the test, the effect might also be further improved.

This article is the author FlyAI platform release of original content, using the Creative Commons Attribution - Noncommercial - No Derivative Works 4.0 International License Agreement for licensing, reproduced, please attach the original source link and this statement.
This link address (direct video to explain): https://www.flyai.com/n/94902