Skip to content

Understanding Results

Face Detection - CSV Output

results.csv

Single File Mode (Image or Video)

  • filename – processed file or video frame name
  • face_detected – 1 if a face was found, 0 otherwise
  • face_count – number of faces in the image or frame
  • face_X_x, face_X_y – center coordinates of each face bounding box
  • face_X_width, face_X_height – dimensions of each face bounding box
  • face_X_confidence – confidence score for each detected face

Folder Mode

Same columns as single file mode but includes entries for every processed file or video frame in the folder.

summary.csv

Single File Mode

  • path – name of the processed file
  • type – image or video
  • total_processed_frames – number of frames processed (1 for images)
  • total_duration – video duration in seconds (N/A for images)
  • processed_frames_with_faces – frames where faces were detected
  • face_percentage – percentage of frames with faces
  • model – detection model used
  • confidence_threshold – threshold applied during detection

Folder Mode

Same columns as single file mode but provides two rows summarising:

  1. All images in the folder
  2. All videos in the folder

Each row contains aggregate values for path, type, total processed frames, total duration, frames with faces, face percentage, model, and confidence threshold.

Image Output

The application saves visual outputs for all images and video frames with detected faces.

Bounding Boxes

  • Green rectangles show each detected face and match the coordinates in results.csv.

Confidence Scores

  • Each bounding box displays the detection confidence between 0 and 1.
  • Values correspond to face_X_confidence in results.csv.

Output File Names

  • For images, the output file retains the original name.
  • For videos, each processed frame is saved separately using the format: [video_name]_[frame_number]_[timestamp].jpg

These outputs make it easy to verify detection results and assess model performance.

"Under Construction" This module is still under active development.

Hand Detection - CSV Output

detections.csv

The 'detections.csv' file contains per detection information, where each record represents one hand detected

  • dataset – Name of processed dataset/Folder title
  • filename – Name of processed image/video frame
  • hand_id – hand identifier assigned to each detected hand within a frame (Starting at 0, i.e if 4 hands are present, labels will be 0,1,2, and 3
  • hand_x1 – x-coordinate of the left edge of the bounding box
  • hand_y1 – y-coordinate of the top edge of the bounding box
  • hand_x2 – x-coordinate of the right edge of the bounding box
  • hand_y2 – y-coordinate of the bottom edge of the bounding box
  • Hand_confidence – Confidence score of the detected hand
  • State – Predicted hand state ID (0 - No touch, 1 - Self touch, 2 - Other touch, 3 - Portable object touch, 4 - Furniture touch)
  • Hand_side – Predicted hand side ('Left' or 'Right')
  • Owner_label – Predicted owner label ('Own' or 'Other')
  • frame_idx – Video frame index, where available
  • state_raw – raw state ID output
  • state_label – human-readable contact state label

summary.csv

The 'summary.csv' file contains a frame-level summary of the overall hand detections. Each record represents one processed image/video frame

  • filename – Name of processed image/video frames
  • frame_idx – Video frame index, where available
  • img_w – image width in pixels
  • img_h – image height in pixels
  • n_hands – Total number of hands detected in the frame
  • n_own – Total number of own hands detected in the frame
  • n_other – Total number of other hands detected in the frame
  • n_state0_none – Total number of detected hands classified as having no contact
  • n_state1_self – Total number of detected hands classified as touching self (Body belonging to hand owner)
  • n_state2_other – Total number of detected hands classified as touching another person (not self)
  • n_state3_portable – Total number of detected hands classified as interacting with a portable object
  • n_state3_furniture – Total number of detected hands classified as interacting with furniture/fixed environmental surface i.e. Floor, door, sofa

Visualisation outputs

Under construction

Baseline

Under construction

Tuned

Under construction

Speech Recognition - CSV Output

detections.csv

Single File Mode (Audio or Video)

  • id – sequential detection/segment ID
  • frame_idx – frame index (blank for audio-only files; used only when detections are tied to video frames)
  • filename – processed audio or video file name
  • mode – detection mode, e.g., speech
  • start, end – start and end time of the segment, in seconds
  • label – detected label (e.g., speech)
  • confidence – model confidence/log-probability score for the segment
  • model – transcription model used (e.g. Whisper (OpenAI))
  • model_size – Whisper model size used (e.g. tiny, base, small, medium, large-v3, turbo), as chosen in the size dropdown
  • text – transcribed text for the segment
  • language – detected or specified language code (e.g. en)
  • speaker – speaker label, if speaker diarization was enabled (blank otherwise)

Folder Mode

Same columns as single file mode but includes entries for every processed audio/video file in the folder, distinguished by filename.

detections_words.csv

Word-level breakdown of each transcribed segment.

  • filename – processed audio or video file name
  • word – individual transcribed word
  • start, end – start and end time of the word, in seconds
  • speaker – speaker label, if diarization was enabled (blank otherwise)
  • word_score – model confidence score for the individual word
  • model – transcription model used (e.g. Whisper (OpenAI))
  • model_size – Whisper model size used (e.g. tiny, base, small, medium, large-v3, turbo), as chosen in the size dropdown
  • segment_start, segment_end – start and end time of the parent segment the word belongs to
  • segment_text – full text of the parent segment, for cross-reference with detections.csv

summary.csv

Single File Mode

  • path – full path of the processed file
  • type – audio or video
  • segments – total number of speech segments detected
  • duration – total duration of the file, in seconds
  • language – detected or specified language code
  • model – transcription model used
  • model_size – Whisper model size used (e.g. tiny, base, small, medium, large-v3, turbo), as chosen in the size dropdown

Folder Mode

Same columns as single file mode but provides one row per processed file.

Per-File Transcription Output

In addition to the combined detections.csv / detections_words.csv, the application also saves an individual pair of CSVs for each processed file, named after the source file:

  • [filename]_transcript.csv – same columns as detections.csv, scoped to that file only
  • [filename]_words.csv – same columns as detections_words.csv, scoped to that file only
  • [filename]_transcript.txt – A human-readable transcript. The first line records the model and size (e.g. # Model: Faster Whisper (small)), followed by one line per detected segment, formatted as:

[start-end] Segment text start, end – segment timestamps in seconds, matching detections.csv

Note

In Single File Mode, these per-file CSVs are identical to detections.csv and detections_words.csv, since only one file is processed. In Folder Mode, they diverge: detections.csv / detections_words.csv aggregate rows across all files in the folder, while [filename]_transcript.csv / [filename]_words.csv contain only the rows for that specific file — useful for reviewing or sharing results on a per-recording basis without filtering the combined output.

Made with Material for MkDocs