SceneClassifier

This repository is for greyscale scene image classification from the in-class Kaggle challenge and NCTU Computer Vision HW.

The dataset is a little different:

• Kaggle challenge: 3859 grey images with 13 categories (train:2819, test:1040)
• CV HW: 1650 grey images with 15 categories (train:1500, test:150)

Previous work

1. VGG16 (imagenet pretrain) + 2*FC layers & Dropout
2. ResNet50 (imagenet pretrain) on Keras 2.2.4 Broken BatchNorm Freeze
3. Image Size: 224, VGG16 preprocess_input + horizontal_flip (on-the-fly data augmentation)
4. Train on spilt training set(loss some of training data)
5. Ensemble prediction on Kaggle 0.899 accuracy

New method

1. ResNet50 (imagenet pretrain) on TF2.2 classification_models
2. CosineAnnealingScheduler
3. Image Size: 256, + horizontal_flip + brightness + zoom + rotation (on-the-fly data augmentation)
4. Train on whole training set
5. Single model prediction on CV HW 0.98 accuracy

Experiment

EfficientNet

ResNet50

Freeze first 12 layers (0~47 layers in the implment)
|Model| Batch_size| Accuracy| Extra|
| ——————— | |:——:|:—-:|
| ResNet50 | 64 | 0.953 |Generate New Data|
| ResNet50 | 64 | 0.966 |on-the-fly|
| ResNet50 | 64 | 0.946 |on-the-fly + constrast_pil|
| ResNet50 | 64 | 0.98 |on-the-fly + rotation 5|
| ResNet50 | 64 | 0.96 |on-the-fly + rotation 7|
| ResNet50 | 64 | 0.953 |on-the-fly + rotation 10|

Big Transfer (BiT)

• BiT-M (pre-trained on ImageNet-21k), on-the-fly

Conclusion

• Use ResNet50 with imagenet pretrain and freeze first 12 layers
• Large batch size might be helpful
• Use on-the-fly (random) instead of generate new data on data augmentation
• Use Brightness, Zoom and Rotation instead of Equalize and RandomResizedCropped
• Use TF2 if you want to freeze BN layers
• Sparse labels might help on accuracy (Dense without softmax, class_mode=’sparse’, loss=SparseCategoricalCrossentropy)