Action recognition basics in Computer Vision - Model Pipeline Trace

Model Pipeline - Action recognition basics

This pipeline learns to recognize human actions from video clips. It processes video frames, extracts important features, trains a model to understand actions, and then predicts the action in new videos.

Data Flow - 5 Stages

1Input Video Frames

100 videos x 30 frames x 64 height x 64 width x 3 channels→Collect raw video frames representing actions→100 videos x 30 frames x 64 height x 64 width x 3 channels

A video showing a person waving, represented as 30 color frames of size 64x64

↓

2Preprocessing

100 videos x 30 frames x 64 height x 64 width x 3 channels→Normalize pixel values to range 0-1 and resize frames if needed→100 videos x 30 frames x 64 height x 64 width x 3 channels

Pixel values changed from 0-255 to 0.0-1.0 for better model training

↓

3Feature Extraction

100 videos x 30 frames x 64 height x 64 width x 3 channels→Use 3D convolution layers to extract motion and appearance features→100 videos x 15 frames x 16 height x 16 width x 64 feature maps

Features capturing movement patterns like waving or jumping

↓

4Temporal Pooling

100 videos x 15 frames x 16 height x 16 width x 64 feature maps→Aggregate features over time to summarize action→100 videos x 64 feature vectors

A single vector representing the whole action in each video

↓

5Classification Model Training

100 videos x 64 feature vectors→Train a neural network classifier to label actions→100 videos x 5 action classes

Labels like 'waving', 'jumping', 'running', 'clapping', 'walking'

Training Trace - Epoch by Epoch

Loss
1.2 |*       
0.9 | *      
0.7 |  *     
0.5 |   *    
0.4 |    *   
    +---------
     1 2 3 4 5 Epochs

Epoch	Loss ↓	Accuracy ↑	Observation
1	1.2	0.40	Model starts learning, accuracy is low
2	0.9	0.55	Loss decreases, accuracy improves
3	0.7	0.68	Model captures action features better
4	0.5	0.78	Good improvement, model generalizes well
5	0.4	0.83	Training converges with high accuracy

Prediction Trace - 5 Layers

Layer 1: Input Video Frames

Layer 2: Preprocessing

Layer 3: 3D Convolutional Feature Extraction

Layer 4: Temporal Pooling

Layer 5: Classification Layer

Model Quiz - 3 Questions

What is the main purpose of the 3D convolutional layer in this pipeline?

A To convert video frames into grayscale images

B To capture motion and appearance features across video frames

C To normalize pixel values between 0 and 1

D To split videos into training and testing sets

Key Insight

Action recognition models learn by extracting motion and appearance features from video frames over time. Normalizing inputs and using 3D convolutions help the model understand actions better. Training shows steady improvement in accuracy as the model learns to distinguish different actions.

Practice

1. What is the main goal of action recognition in computer vision?

easy

A. To generate captions for images

B. To detect objects in images

C. To enhance image resolution

D. To identify human movements in videos

Solution

1. Step 1: Understand the purpose of action recognition

   Action recognition focuses on understanding what actions or movements humans perform in videos.

2. Step 2: Compare with other tasks

   Detecting objects, generating captions, or enhancing resolution are different tasks unrelated to recognizing actions.

3. Final Answer:

   To identify human movements in videos -> Option D

4. Quick Check:

   Action recognition = Identify human movements [OK]

Hint: Action recognition = understanding human movements in videos [OK]

Common Mistakes:

• Confusing action recognition with object detection
• Thinking it generates image captions
• Assuming it improves image quality

2. Which of the following is the correct way to represent a video input for an action recognition model?

easy

A. A sequence of image frames

B. A single grayscale image

C. A text description of the action

D. A 1D audio signal

Solution

1. Step 1: Identify video data format

   Videos are made of many image frames shown in order, so a sequence of frames is the correct input.

2. Step 2: Eliminate incorrect options

   A single image or text or audio does not represent the full video needed for action recognition.

3. Final Answer:

   A sequence of image frames -> Option A

4. Quick Check:

   Video input = sequence of frames [OK]

Hint: Videos = many frames in order, not single images [OK]

Common Mistakes:

• Using a single image instead of multiple frames
• Confusing video input with text or audio
• Ignoring the temporal sequence of frames

3. Consider this Python snippet for extracting features from video frames for action recognition:

features = []
for frame in video_frames:
    feat = extract_features(frame)
    features.append(feat)
print(len(features))

If video_frames contains 10 frames, what will be the output?

medium

A. 10

B. 9

C. 0

D. Error

Solution

1. Step 1: Understand the loop over frames

   The loop runs once for each frame in video_frames, which has 10 frames.

2. Step 2: Count how many features are appended

   Each iteration appends one feature, so after 10 iterations, features has length 10.

3. Final Answer:

   10 -> Option A

4. Quick Check:

   Number of frames = features length = 10 [OK]

Hint: One feature per frame means length equals number of frames [OK]

Common Mistakes:

• Off-by-one errors counting features
• Assuming extract_features returns multiple items
• Thinking the list is empty before print

4. You have this code snippet for action recognition training:

for video, label in dataset:
    features = extract_features(video)
    prediction = model.predict(features)
    loss = loss_function(prediction, label)
    optimizer.zero_grad()
    loss.backward()
    optimizer.step()

The training loss does not decrease after many epochs. What is a likely error?

medium

A. Optimizer step is missing

B. Loss function is not called

C. Features are extracted frame-by-frame but model expects video clips

D. Labels are not used in prediction

Solution

1. Step 1: Analyze feature extraction and model input

   If features are extracted frame-by-frame but the model expects a clip (multiple frames together), the input shape mismatch can cause poor learning.

2. Step 2: Check other training steps

   Loss function is called, optimizer steps are present, and labels are used in loss, so these are correct.

3. Final Answer:

   Features are extracted frame-by-frame but model expects video clips -> Option C

4. Quick Check:

   Input shape mismatch = training loss stuck [OK]

Hint: Check if model input matches feature extraction format [OK]

Common Mistakes:

• Ignoring input shape mismatch
• Assuming loss or optimizer calls are missing
• Not verifying label usage in loss

5. You want to improve an action recognition model that uses only spatial features from single frames. Which approach is best to capture motion information?

hard

A. Train on grayscale frames instead of color

B. Use 3D convolutional neural networks on video clips

C. Add dropout layers to the model

D. Increase image resolution of single frames

Solution

1. Step 1: Understand spatial vs temporal features

   Spatial features come from single frames; motion requires temporal features across frames.

2. Step 2: Identify model type capturing motion

   3D CNNs process multiple frames together, capturing motion and temporal info effectively.

3. Step 3: Evaluate other options

   Increasing resolution, dropout, or grayscale do not add motion info.

4. Final Answer:

   Use 3D convolutional neural networks on video clips -> Option B

5. Quick Check:

   3D CNNs capture motion = better action recognition [OK]

Hint: Motion needs temporal models like 3D CNNs, not just images [OK]

Common Mistakes:

• Thinking higher resolution adds motion info
• Confusing regular CNNs with 3D CNNs
• Ignoring temporal dimension in videos