camarcano/glove_tracking_MLB_april_2024

Glove Tracking Dataset Description

1. Dataset Overview and Scope

This dataset originates from the glove_tracker.py script, which utilizes YOLO (You Only Look Once) object detection models to analyze baseball game videos. The primary goal is to track the positions of key objects frame-by-frame: the catcher's glove, homeplate, and the baseball. This detailed tracking allows for quantitative analysis of catcher movements, positioning, and potentially pitcher command when combined with other data sources.

2. Key Aspects of the Data

• Frame-by-Frame Tracking: The data provides positional information for each processed frame of the source video, creating a time-series representation of object movements.
• Object Detection: It includes raw detection results from the YOLO model, such as the pixel coordinates of the center of detected objects (glove, homeplate, baseball) and the confidence score of each detection. Confidence scores indicate the model's certainty about the detection's accuracy.
• Coordinate Transformation: The script calculates real-world coordinates (in inches) for the glove and baseball relative to the center of the detected homeplate. This transformation relies on establishing a scale factor (pixels_per_inch) based on the known width of homeplate (17 inches). The real-world coordinate system typically uses the center of homeplate as the origin (0,0), with the X-axis horizontal (positive towards the catcher's left/batter's right, matching the standard Statcast view from behind the catcher) and the Y-axis vertical (positive upwards). Accuracy is contingent on a clear, undistorted view of homeplate in the frame used for reference.
• Glove Path Processing: It includes columns (glove_processed_x, glove_processed_y) representing a potentially smoothed or interpolated path of the catcher's glove. This processing, handled by functions like _fill_missing_detections and _handle_missing_detections in the script, aims to create a continuous trajectory, filling gaps where the YOLO model might have momentarily failed to detect the glove due to occlusion or motion blur. The is_interpolated column explicitly indicates whether the processed coordinates for a given frame are based on a direct detection (False) or were estimated (True). Velocity-based filtering (_filter_glove_detection) is also applied to remove outlier detections that represent physically impossible movements.
• Metadata: Includes the frame_idx (frame number) and the original video_filename for reference and traceability back to the source video. The presence of gamepk and playid suggests these were added subsequently (likely based on the video_filename or external mapping) to link the tracking data to specific MLB games and plays.

3. Coordinate System and Transformation Details

Understanding the coordinate systems is crucial for accurate analysis: Pixel Coordinates (_center_x, _center_y): These are raw outputs from the object detection model, relative to the top-left corner (0,0) of the video frame. They are useful for debugging detections but not directly comparable across videos or for real-world measurements. Real-World Coordinates (_real_x, _real_y): This is the primary coordinate system for analysis.

• Origin: The detected center of homeplate.
• Units: Inches.
• X-axis: Horizontal. Positive values are to the catcher's left (first-base side for a right-handed batter), negative values are to the catcher's right (third-base side for a right-handed batter). This aligns with the standard Statcast coordinate system.
• Y-axis: Vertical. Positive values are upwards from the center of homeplate, negative values are downwards.
• Calculation: Derived by:
  • Finding the difference in pixels between the object's center and the homeplate's center.
  • Dividing by pixels_per_inch (calculated from the detected pixel width of homeplate, assuming it's 17 inches wide).
  • Inverting the Y-axis pixel difference because pixel Y increases downwards, while real-world Y increases upwards. Processed Coordinates (_processed_x, _processed_y): These share the same real-world coordinate system (inches, relative to homeplate center) as _real_x and _real_y, but represent the cleaned-up path after interpolation and filtering have been applied. These are generally the best coordinates to use for analyzing the glove's continuous movement trajectory.

4. Glove Path Processing Explained

The glove_processed_x and glove_processed_y columns provide a more robust representation of the glove's path than the raw glove_real_x/y data, which can contain gaps. Interpolation (_fill_missing_detections, _handle_missing_detections): When the YOLO model fails to detect the glove in a frame (or the detection is filtered out), the script attempts to fill the gap. Simple methods might involve linear interpolation between the last known good position and the next known good position. More advanced handling might connect sequences if the gap isn't too large (e.g., < 10 frames). Filtering (_filter_glove_detection): Before coordinates are considered "valid" (and used for interpolation or processing), they might be filtered based on physical plausibility. The script includes velocity-based filtering, discarding detections that would imply the glove moved faster than a defined threshold (e.g., max_velocity_inches_per_sec = 120.0) between frames. This helps remove erroneous detections that jump unnaturally. is_interpolated Flag: This flag is essential for understanding the reliability of the glove_processed coordinates at any given frame. False indicates high confidence (based on a direct, filtered detection), while True indicates the position was estimated. Analyses focusing on precise positioning (like identifying the target frame) should prioritize frames where is_interpolated is False.

5. Linking to Baseball Savant/MLB Data

The video_filename helps identify the source video. If this filename, or the added gamepk and playid columns, contains the correct identifiers: gamepk: The official MLB identifier for the game. Can be used to query the MLB Stats API or find the game on sites like Baseball Savant. It provides context like teams, date, and venue. playid: The unique Statcast identifier for a specific pitch or play event within a game. This is the most direct way to link a sequence of rows corresponding to a pitch (identified by frame_idx range) to the specific event on Baseball Savant or in Statcast data downloads. This allows merging glove tracking data with detailed pitch outcomes, trajectories, and classifications (e.g., pitch type, velocity, location, result).

6. Potential Use Cases and Analyses

This dataset is valuable for analyzing catcher setup, glove movement patterns, and potentially comparing target locations with actual pitch results when combined with Statcast data using the gamepk and playid. Specific analyses could include:

• Catcher Setup Analysis: Analyzing glove_processed_x and glove_processed_y in frames just before the pitch release to quantify setup location consistency across different pitch types, counts, or batters.
• Pre-Pitch Movement: Measuring the timing and extent of glove movement from the initial setup position to the final target presentation frame (intent_frame if calculated by command_analyzer.py using this data).
• Target vs. Actual (Command Analysis): By merging this data with Statcast plate_x and plate_z using gamepk and playid, one can calculate the deviation between the intended glove target (using glove_processed_x/y from an identified intent frame) and the actual pitch location. This requires careful identification of the correct "intent frame" prior to pitch release. (Note: This deviation calculation is the primary purpose of the separate command_analyzer.py script).
• Glove Path Smoothness/Efficiency: Analyzing the velocity and acceleration derived from the glove_processed_x/y coordinates to quantify how smoothly or efficiently the catcher moved their glove, potentially correlating with framing ability or pitch type.
• Heatmaps and Visualizations: Using the glove_processed_x/y data to generate heatmaps (as done by plot_glove_heatmap in the script) showing common setup locations or areas of frequent glove movement for different situations.

7. Data Limitations and Considerations

Model Accuracy: The accuracy of all coordinates depends on the performance of the underlying YOLO object detection model. Errors in detection (incorrect bounding boxes, misclassifications) will propagate into the coordinate calculations. Camera Angle and Distortion: The transformation from pixels to inches assumes a relatively stable camera view perpendicular to the plane of interest and minimal lens distortion. Significant changes in camera angle or severe distortion can impact the accuracy of the pixels_per_inch calculation and subsequent real-world coordinates. Occlusion: The glove or baseball can be occluded by the batter, umpire, or other objects, leading to missed detections and reliance on interpolation. homeplate Detection: The entire real-world coordinate system relies on accurate and consistent detection of homeplate to establish the origin (0,0) and the scale (pixels_per_inch). If homeplate detection fails or is inconsistent, the calculated _real_x, _real_y, and _processed_x/y coordinates will be inaccurate. Interpolation Artifacts: While interpolation fills gaps, it's an estimation. The is_interpolated flag should be used to identify these estimated points, especially for analyses requiring high positional accuracy.

8. Dataset Dictionary

Column Name	Description	Source / Calculation
frame_idx	The index number of the video frame (starting from 0).	Video Processing (glove_tracker.py)
homeplate_center_x	The horizontal pixel coordinate of the detected center of the homeplate in the frame.	YOLO Detection (glove_tracker.py)
homeplate_center_y	The vertical pixel coordinate of the detected center of the homeplate in the frame.	YOLO Detection (glove_tracker.py)
homeplate_width	The width of the detected homeplate bounding box in pixels. Used to calculate pixels_per_inch.	YOLO Detection (glove_tracker.py)
homeplate_confidence	The confidence score (0.0 to 1.0) from the YOLO model for the homeplate detection.	YOLO Detection (glove_tracker.py)
glove_center_x	The horizontal pixel coordinate of the detected center of the catcher's glove in the frame.	YOLO Detection (glove_tracker.py)
glove_center_y	The vertical pixel coordinate of the detected center of the catcher's glove in the frame.	YOLO Detection (glove_tracker.py)
glove_confidence	The confidence score (0.0 to 1.0) from the YOLO model for the glove detection.	YOLO Detection (glove_tracker.py)
glove_real_x	The calculated horizontal position of the glove (inches) relative to the homeplate center (X=0). Negative is catcher's right, positive is left.	Calculation (glove_tracker.py)
glove_real_y	The calculated vertical position of the glove (inches) relative to the homeplate center (Y=0). Positive is up, negative is down.	Calculation (glove_tracker.py)
baseball_center_x	The horizontal pixel coordinate of the detected center of the baseball in the frame.	YOLO Detection (glove_tracker.py)
baseball_center_y	The vertical pixel coordinate of the detected center of the baseball in the frame.	YOLO Detection (glove_tracker.py)
baseball_confidence	The confidence score (0.0 to 1.0) from the YOLO model for the baseball detection.	YOLO Detection (glove_tracker.py)
baseball_real_x	The calculated horizontal position of the baseball (inches) relative to the homeplate center (X=0).	Calculation (glove_tracker.py)
baseball_real_y	The calculated vertical position of the baseball (inches) relative to the homeplate center (Y=0).	Calculation (glove_tracker.py)
pixels_per_inch	The calculated scale factor (pixels / inch) based on the detected homeplate width (assuming 17 inches). Used for coordinate conversion.	Calculation (glove_tracker.py)
glove_processed_x	The horizontal position of the glove (inches, relative to homeplate center) after processing (interpolation/filtering) to create a continuous path.	Calculation (glove_tracker.py)
glove_processed_y	The vertical position of the glove (inches, relative to homeplate center) after processing (interpolation/filtering) to create a continuous path.	Calculation (glove_tracker.py)
is_interpolated	1 if glove_processed coords for this frame were estimated/interpolated; 0 if based on a direct detection.	Calculation (glove_tracker.py)
video_filename	The basename of the source video file processed.	Metadata (glove_tracker.py)
gamepk	MLB Game ID.	External / Metadata / Statcast
playid	Statcast Play ID.	External / Metadata / Statcast