The Wowza Video Intelligence Framework generates structured JSON log entries for the media analyzed by the Video Intelligence Service (VIS). Each entry records detection results, processing‑performance metrics, and tracking data in a machine‑readable format designed for enterprise observability and SIEM platforms.
Prerequisite
VIF file logging requires the LogFiles event listener to be enabled. If it is not, the wowzastreamingengine_vi_[date].log file described below is not written.
VIF logs are written to the WSE logs directory alongside the standard engine logs:
[install-dir]/logs/wowzastreamingengine_vi_[date].log
Each line is a self‑contained JSON object, so VIF logs are directly ingestible by Splunk, Elastic, Datadog, and similar tools without extra parsing.
VIF generates logs from two sources:
- VIC logs (inside WSE), written to the standard WSE log files (
wowzastreamingengine_access.log,…_error.log, and…_vi.logwhen LogFileEvent is enabled). Available in./wse/logs/when volume mounts are enabled. - VIS logs, written by the VIS container. View with
docker compose --profile vi-service logs -f.
On‑demand (VOD) results. For file jobs, per‑window results are stored as JSONL and downloadable from the On‑Demand Jobs list (the JSONL action) or the results viewer. On disk they live under the configured
jobs_dir(defaultvif-vod-jobs/).
Debug timing output. When Log Timing is set on a stream config, VIF writes timing/diagnostic detail to
/tmp/vif/<stream_name>at the configured interval (0 disables). This is separate from the LogFiles listener log.
Log structure
Every VIF log entry is a JSON object with two top‑level properties:
time_stamp, server‑local timestamp when the entry was written to disk.vif_data, all detection results, performance metrics, and metadata for the analyzed sample.
All fields below are properties of vif_data unless noted.
Root‑level fields
| Field | Description |
|---|---|
time_stamp |
Date/time the entry was written (YYYY‑MM‑DD HH:mm:ss, server local). |
type |
Log event category. Currently detections. |
detector_type |
The detector that produced the result: object, scene, motion, vlm, synthetic, or chain. [Updated: adds motion, chain] |
stream_name |
The WSE application stream being analyzed. |
request_id |
UUID v4 for the inference request; correlate with webhook payloads and VIS diagnostics. |
created_at |
ISO‑8601 UTC timestamp when VIS produced the result. |
detection_window |
Frame‑range metadata for the analyzed sample. See below. |
preprocess_time_ms / inference_time_ms / postprocess_time_ms / total_processing_time_ms |
VIS timing metrics (see Processing performance). |
detections_type |
Classifies the results in this entry. In addition to object, it writes values for the new detector types (motion, chain, and the other detectors). The exact strings are confirmed in the plugin Readme. |
sequence_id |
Monotonically increasing integer ordering results per stream. |
| result payload | The payload varies by detector. Object, scene, and motion detectors carry detections; VLM, synthetic, and chain detectors carry their own result shapes. Omitted when empty. |
Detection window fields
The detection_window object describes the analyzed frame range.
Note: Window‑based detectors (scene, VLM, synthetic, and chains running at window cadence) analyze a span of frames, so
from_frame_id,to_frame_id, andframe_countdescribe a real range. Frame‑cadence detectors (object, motion) still report a single frame (frame_count1, equal from/to).
| Field | Description |
|---|---|
from_frame_id / to_frame_id |
First/last frame IDs in the window. |
frame_count |
Number of frames in the window. |
from_frame / to_frame |
Presentation timestamps of the first/last frames (YYYY‑MM‑DD HH:mm:ss.SSS). |
When VIF cannot place a window in the stream’s time base, it omits the four time fields (from_frame, to_frame, from_time_code, to_time_code) from the entry rather than writing a 1970 date, and behind_live_ms is present only when it can be resolved. Treat these fields as optional when you parse webhooks, ID3 tags, logs, or REST results.
Detection object fields (object / motion)
| Field | Description |
|---|---|
class_name |
For object: the COCO class (e.g. person, car). For motion: always motion. |
confidence |
For object: model certainty (0.0–1.0). For motion: the track’s net travel as a fraction of the frame diagonal, not a quality score. [Updated: motion meaning] |
track_id |
Instance identifier for cross‑frame tracking. |
frame_id |
Frame in which the detection occurred. |
bbox |
Bounding box (x, y, x2, y2, w, h, in pixels from the top‑left). For motion, the box is the span of movement, not the subject’s outline. |
VLM entries carry vlm (content/reasoning or a custom schema), synthetic entries carry synthetic (verdict + score), and chain entries carry rings with each stage’s typed payload. See the respective detector guides.
Supported object classes
The default RF‑DETR model included with VIF is pre‑trained on COCO and detects 80 classes (person, bicycle, car, …, toothbrush). Custom/BYOM models produce the labels defined in their training configuration.
Processing performance metrics
| Metric | Typical range | Description |
|---|---|---|
preprocess_time_ms |
0–2 ms | Frame preparation (decode, resize, normalize). |
inference_time_ms |
6–20 ms | Model execution on GPU. |
postprocess_time_ms |
0–5 ms | Suppression, thresholding, tracking. |
total_processing_time_ms |
10–40 ms | Total VIS pipeline time. |
Times rise with object count, model complexity, or GPU load. Window‑based and chained detectors cost more per result than single‑frame object detection.
Sample log entries
Frame with no detections
{"time_stamp":"2026-02-28 00:39:08","vif_data":{
"type":"detections","detector_type":"object","stream_name":"lobby-east",
"request_id":"1e33e303-77ee-40ca-8708-56d4b69bcde8","created_at":"2026-02-28T00:39:07Z",
"detection_window":{"from_frame_id":0,"to_frame_id":0,"frame_count":1,
"from_frame":"2026-02-28 00:39:07.000","to_frame":"2026-02-28 00:39:07.000"},
"preprocess_time_ms":1,"inference_time_ms":8,"postprocess_time_ms":1,
"total_processing_time_ms":20,"detections_type":"object","sequence_id":0}}
Frame with multiple detections
{"time_stamp":"2026-02-28 04:40:24","vif_data":{
"type":"detections","detector_type":"object","stream_name":"parking-lot",
"request_id":"cdb46f9f-b2e6-497e-8104-d3e132b50bb0","created_at":"2026-02-28T04:40:22Z",
"detection_window":{"...":"..."},
"preprocess_time_ms":0,"inference_time_ms":6,"postprocess_time_ms":1,
"total_processing_time_ms":13,"detections_type":"object","sequence_id":1446,
"detections":[
{"class_name":"person","confidence":0.94,"track_id":72,"frame_id":2788,
"bbox":{"x":366,"y":85,"x2":456,"y2":292,"w":90,"h":207}},
{"class_name":"tv","confidence":0.87,"track_id":71,"frame_id":2788,
"bbox":{"x":153,"y":176,"x2":233,"y2":225,"w":80,"h":49}}]}}
For the motion, synthetic, VLM, and chain result shapes, see Use motion detection in VIF, Use the Synthetic Video Detector in VIF, Use VLM analysis in VIF, and Use model chaining in VIF.
Log output channels
VIF detection data is routed simultaneously to independent channels; the filesystem log in this article is one of them:
| Channel | Description |
|---|---|
| Filesystem logs | Structured JSON written locally for Splunk/Elastic/Datadog, etc. |
| Enriched HLS stream | ID3 metadata with detection JSON injected into HLS. |
| Visual overlays | Boxes and labels burned into the video frame. |
| Webhook dispatch | HTTP POST per detection event to configured endpoints. |
| Custom Java listeners | WSE listener interface for custom processing. |
| On‑demand JSONL results | Per‑window results stored for file jobs. |




