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

Model	Batch_size	Accuracy	Extra
EfficientNetB0	64	0.92
EfficientNetB0	64	0.906	noisy-student pretrain
EfficientNetB1	64	0.926
EfficientNetB1	64	0.906	noisy-student pretrain
EfficientNetB4	16	0.92
EfficientNetB4	32	0.95
EfficientNetB4	32	0.89	Freeze 1st Block(Conv+BN+Activation)
EfficientNetB4	32	0.9	Freeze 1~2 Blocks(Conv+BN+Activation)
EfficientNetB5	16	0.926
EfficientNetB6	16	0.9	Freeze 1st Block(Conv+BN+Activation)
EfficientNetB6	16	0.926	Freeze 1~2 Blocks(Conv+BN+Activation)
EfficientNetB6	16	0.94	Freeze 1~3 Blocks(Conv+BN+Activation)
EfficientNetB6	16	0.85	Freeze 1~4 Blocks(Conv+BN+Activation)

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

Model	Batch_size	Accuracy	Extra
R50x1	64	0.966
R50x3	64	0.96
R101x1	64	0.96
R101x3	64	0.953

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)