Object tracking basics in Computer Vision - Model Metrics & Evaluation

Metrics & Evaluation - Object tracking basics

Which metric matters for Object tracking basics and WHY

In object tracking, we want to know how well the model follows the object over time. Key metrics are Multiple Object Tracking Accuracy (MOTA) and Multiple Object Tracking Precision (MOTP). MOTA measures how many mistakes the tracker makes, like missing objects or wrong matches. MOTP measures how close the predicted object positions are to the true positions. These metrics help us understand both the correctness and the precision of tracking.

Confusion matrix or equivalent visualization

    Tracking Confusion Matrix (simplified):

          | Predicted Object Present | Predicted Object Absent |
    ------|--------------------------|-------------------------|
    True  | True Positive (TP)        | False Negative (FN)      |
    False | False Positive (FP)       | True Negative (TN)       |

    TP: Correctly tracked objects
    FP: Tracker reports object but none exists
    FN: Tracker misses an object

    Total objects = TP + FN
    Total predictions = TP + FP
    

Precision vs Recall tradeoff with concrete examples

Precision tells us how many tracked objects are actually correct. High precision means few false alarms.

Recall tells us how many true objects were found by the tracker. High recall means few missed objects.

Example: In a security camera, if the tracker has high precision but low recall, it rarely mistakes background for people but misses some people. If it has high recall but low precision, it finds almost all people but sometimes thinks shadows are people.

Balancing precision and recall depends on the use case. For safety, high recall is often more important to avoid missing objects.

What "good" vs "bad" metric values look like for object tracking

Good tracking:

• MOTA close to 1.0 (near 100%) means very few errors.
• MOTP low error distance (small pixel difference) means precise location.
• High precision and recall (both above 0.8) means tracker finds most objects and rarely makes mistakes.

Bad tracking:

• MOTA below 0.5 means many missed or wrong tracks.
• MOTP high error distance means poor localization.
• Low precision means many false alarms; low recall means many missed objects.

Metrics pitfalls in object tracking

• Ignoring ID switches: Tracker may confuse object identities, which hurts tracking quality but may not show in simple accuracy.
• Overfitting to training videos: Tracker works well on known scenes but fails in new environments.
• Data leakage: Using future frames or ground truth in training can inflate metrics unfairly.
• Accuracy paradox: High accuracy can be misleading if many frames have no objects (tracker predicts no object and is 'correct').

Self-check question

Your object tracker has 98% accuracy but only 12% recall on objects. Is it good for production? Why or why not?

Answer: No, it is not good. The high accuracy likely comes from many frames without objects, so predicting no object is often correct. But 12% recall means the tracker misses 88% of objects, which is poor for tracking. The tracker fails to find most objects, so it is unreliable.

Key Result

MOTA and MOTP are key metrics showing tracking correctness and precision; balance precision and recall to avoid missing or falsely detecting objects.

Practice

1. What is the main goal of object tracking in computer vision?

easy

A. To follow the position of an object across video frames

B. To classify objects into categories

C. To enhance the resolution of video frames

D. To detect edges in a single image

Solution

1. Step 1: Understand object tracking purpose

   Object tracking aims to keep track of where an object moves in a video over time.

2. Step 2: Differentiate from other tasks

   Classification sorts objects into groups, enhancement improves image quality, and edge detection finds boundaries in one image, not tracking movement.

3. Final Answer:

   To follow the position of an object across video frames -> Option A

4. Quick Check:

   Object tracking = follow object position [OK]

Hint: Tracking means following object movement frame by frame [OK]

Common Mistakes:

• Confusing tracking with classification
• Thinking tracking enhances image quality
• Mixing tracking with edge detection

2. Which of the following is the correct way to initialize an OpenCV tracker in Python?

easy

A. tracker = cv2.TrackerKCF()

B. tracker = cv2.Tracker_create('KCF')

C. tracker = cv2.createTracker('KCF')

D. tracker = cv2.TrackerKCF_create()

Solution

1. Step 1: Recall OpenCV tracker initialization syntax

   In OpenCV, trackers like KCF are created using the method cv2.TrackerKCF_create().

2. Step 2: Check each option

   tracker = cv2.Tracker_create('KCF') uses a non-existent method, C uses wrong method name, D tries to instantiate class directly which is incorrect.

3. Final Answer:

   tracker = cv2.TrackerKCF_create() -> Option D

4. Quick Check:

   OpenCV tracker init = TrackerKCF_create() [OK]

Hint: OpenCV trackers use create() methods, not constructors [OK]

Common Mistakes:

• Using wrong method names
• Trying to instantiate tracker classes directly
• Confusing createTracker with Tracker_create

3. Given the following Python code snippet using OpenCV tracker, what will be printed?

import cv2
tracker = cv2.TrackerKCF_create()
frame = cv2.imread('frame1.jpg')
bbox = (100, 100, 50, 50)
tracker.init(frame, bbox)
success, box = tracker.update(frame)
print(success, box)

medium

A. True and the original bbox tuple

B. True and a tuple with updated bounding box coordinates

C. SyntaxError due to wrong method names

D. False and None

Solution

1. Step 1: Understand tracker initialization and update

   The tracker is initialized with a frame and bbox, then updated on the same frame.

2. Step 2: Predict update output on same frame

   Since the frame hasn't changed, the tracker will succeed and return the updated bbox coordinates (which may be the same as original but returned as a new tuple).

3. Final Answer:

   True and a tuple with updated bounding box coordinates -> Option B

4. Quick Check:

   Update on same frame = success True, bbox updated [OK]

Hint: Update on same frame returns success True and updated box [OK]

Common Mistakes:

• Expecting bbox to change without new frame
• Confusing success flag meaning
• Assuming syntax errors in correct code

4. Identify the error in this object tracking code snippet:

import cv2
tracker = cv2.TrackerKCF_create()
frame = cv2.imread('frame1.jpg')
bbox = (100, 100, 50, 50)
tracker.init(frame, bbox)
success, box = tracker.update()
print(success, box)

medium

A. Missing frame argument in tracker.update() call

B. bbox should be a list, not a tuple

C. tracker.init() should be called after update()

D. cv2.imread() returns None if file not found

Solution

1. Step 1: Check tracker.update() method signature

   The update() method requires the current frame as an argument to track the object.

2. Step 2: Identify missing argument

   The code calls update() without any argument, causing an error.

3. Final Answer:

   Missing frame argument in tracker.update() call -> Option A

4. Quick Check:

   update() needs frame argument [OK]

Hint: update() always needs the current frame as input [OK]

Common Mistakes:

• Forgetting frame argument in update()
• Thinking bbox must be list not tuple
• Calling init() after update()

5. You want to track a fast-moving object in a video with limited computing power. Which tracker should you choose for a good balance of speed and accuracy?

hard

A. MIL tracker

B. CSRT tracker

C. KCF tracker

D. BOOSTING tracker

Solution

1. Step 1: Understand tracker characteristics

   KCF tracker is known for good speed and decent accuracy, suitable for fast objects and limited resources.

2. Step 2: Compare with other trackers

   MIL and BOOSTING are slower or less accurate; CSRT is more accurate but slower and heavier on computation.

3. Final Answer:

   KCF tracker -> Option C

4. Quick Check:

   Fast object + limited power = KCF tracker [OK]

Hint: KCF balances speed and accuracy well for fast objects [OK]

Common Mistakes:

• Choosing CSRT for speed instead of accuracy
• Assuming BOOSTING is fastest
• Ignoring resource limits