Fine-tuning SOTA video models on your own dataset - Sign Language

Question

I am trying to implement a sign classifier using gluoncv API as part of my final year college project.

Data set: http://facundoq.github.io/datasets/lsa64/

I followed the Fine-tuning SOTA video models on your own dataset tutorial and fine-tuned. Tutorial: https://cv.gluon.ai/build/examples_action_recognition/finetune_custom.html

1. i3d_resnet50_v1_custom Accuracy Graph I3D

2. slowfast_4x16_resnet50_custom Accuracy Graph Slow Fast

The plotted graph showing almost 90% accuracy but when I running on my inference I am getting miss classification even on the videos I used to train.

So I am stuck, could you have some guide to give anything will be help full.

Thank you

My data loader for I3D:

num_gpus = 1
ctx = [mx.gpu(i) for i in range(num_gpus)]
transform_train = video.VideoGroupTrainTransform(size=(224, 224), scale_ratios=[1.0, 0.8], mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
per_device_batch_size = 5
num_workers = 0
batch_size = per_device_batch_size * num_gpus

train_dataset = VideoClsCustom(root=os.path.expanduser('DataSet/train/'),
                               setting=os.path.expanduser('DataSet/train/train.txt'),
                               train=True,
                               new_length=64,
                               new_step=2,
                               video_loader=True,
                               use_decord=True,
                               transform=transform_train)

print('Load %d training samples.' % len(train_dataset))
train_data = gluon.data.DataLoader(train_dataset, batch_size=batch_size,
                                   shuffle=True, num_workers=num_workers)

Inference running:

from gluoncv.utils.filesystem import try_import_decord
decord = try_import_decord()

video_fname = 'DataSet/test/006_001_001.mp4'
vr = decord.VideoReader(video_fname)
frame_id_list = range(0, 64, 2)
video_data = vr.get_batch(frame_id_list).asnumpy()
clip_input = [video_data[vid, :, :, :] for vid, _ in enumerate(frame_id_list)]

transform_fn = video.VideoGroupValTransform(size=(224, 224), mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
clip_input = transform_fn(clip_input)
clip_input = np.stack(clip_input, axis=0)
clip_input = clip_input.reshape((-1,) + (32, 3, 224, 224))
clip_input = np.transpose(clip_input, (0, 2, 1, 3, 4))
print('Video data is readed and preprocessed.')

# Running the prediction
pred = net(nd.array(clip_input,  ctx = mx.gpu(0)))
topK = 5
ind = nd.topk(pred, k=topK)[0].astype('int')
print('The input video clip is classified to be')
for i in range(topK):
    print('\t[%s], with probability %.3f.'%
          (CLASS_MAP[ind[i].asscalar()], nd.softmax(pred)[0][ind[i]].asscalar()))

Answer 1

I found my bug, this is happening because of less augmentation so I changed my transform on both train data loader and inference like below now its working properly.

transform_train = transforms.Compose([
    # Fix the input video frames size as 256×340 and randomly sample the cropping width and height from
    # {256,224,192,168}. After that, resize the cropped regions to 224 × 224.
    video.VideoMultiScaleCrop(size=(224, 224), scale_ratios=[1.0, 0.875, 0.75, 0.66]),
    # Randomly flip the video frames horizontally
    video.VideoRandomHorizontalFlip(),
    # Transpose the video frames from height*width*num_channels to num_channels*height*width
    # and map values from [0, 255] to [0,1]
    video.VideoToTensor(),
    # Normalize the video frames with mean and standard deviation calculated across all images
    video.VideoNormalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])