Changelogs
Hello, in this part we introduce a new way for us to explain to you what we've improved/changed. Please use these changelogs to monitor how we improve our products!
- Add NFV4-FR filtering based on bounding box, including new parameters:
bbox_filter_active
,bbox_filter_threshold
- Add new analytic people counting based on head with
NFV4-HC
- Multiple MJPEG server port support
- Expose NFV4-FR parameters
frontal_only
,matching_iou_threshold
to adjust recognition sensitivity - Add MP1-Streamer support
- New analytic NFV4-GOC with purpose of General Object Counting consist with 80 class. (POC Support only)
- Streams are instantiated as subprocess
- Add stream_child support
- Add direction detection for NFV4-LPR2
- Add whitelist/blacklist option ROI for NFV4-LPR2
- Update CE new model with improvement in pipeline to reduce latency by 100ms
- Solving case in
NFV4-FR
where face that too close to camera is not dumped Solve false positive case in People Counting
- New url on
/mjpeg
endpoint with query params?fps=xx
,?height=xx
, and?jq=xx
- Enable dashboard stream visual setting with parameter
height
andjpeg_quality
- Set global configuration on jpeg quality image dump each analytic_id with default analytic_id FR jq=80, otherwise jq=60
- Disable
std::terminate
if pipeline is unresponsive
- Fix False Negative on
NFV4-PPE
- New model of
NFV4-CE
and use counting points as visualization
- New pipeline
NFV4-PPE
that provides PPE (Personal Protective Equipment) detection
- New pipeline
NFV4-VD
that provides vehicle dwelling information
- New pipeline
NFV4-VCAT
to use vehicle recognition (enrollment based) in vehicle counting
- Introduced a new version of CE : Low Density Crowd Estimation (CELD) which using similar principle with people dwelling.
- Add method pause and resume abstraction
- Pause frame if it isn't invoking
get_active_frame()
- Add
check_capability
on method PUT
- Add new analytics LPRS - LPR for special case
- Assign blacklisted area during assign analytic for FR analytic
- Concatenate camera metrics to v4 metrics
- add log camera routine and print camera logge
- Embedding information now available on event as vector of float
- Embedding on event can be enabled by adding
--fr-embedding-event-enable
on runtime argument
- Add Face Recognition (FR) threshold to control FR dump
- Update pipeline for adding FR events age in each track_id
- Fix crowd estimation negative estimate
- Update person detection model to improve performance and accuracy.
- new analytic Hybrid Face Recognition (NFV4H-FR)
- Add new analytic data to v4bench
- Add synchronous checking capability at create stream by adding key
check_capability
equal totrue
- Update model for VC to improve our performance and accuracy
- Vehicle Aerial (NFV4-VAS) fix stuck when detection is empty
- new analytic Vehicle Aerial Surveillance (NFV4-VAS)
- Integrate ability camera (Face detection code:
NFV4-FR-H
) through websocket to face recognition pipeline. - Transform secondary face landmark value.
- FR interval dumping for each camera assignment using payload, with default value of 0
- Update VC-HW detection model to use stream batching TensorRT with fp16 enabled
- Update pipeline fraction for VC, PC, PD, FR, LPR2, and WLM for A2 and A30
- Test driver face recognition
- Fixing switch between primary and secondary face detection
- Update crowd model to general, improve high crowd counting.
- FR add TensorRT stream batching and fp16 implementation for primary FD (not used)
- FR fix secondary FD landmark output
- add blacklist ROI config to LPR2
- add remove bbox inside blacklist ROI
- Fix LPR2 for T4 scaling
- Internal updates
- Civilian area code for plate regex in LPR2
- Internal minor update.
- We have optimized our LPRv2 Performance by increasing the number of maximum pipelines from 4 to 9 pipelines.
- fix People Dwelling service.
- We add new analytic : People Dwelling (NFV4-PD)
- We changed Person Detection and use a new batching mechanism and enabled FP16 which increased our performance from 10 pipelines to 37 pipelines (increasing 270% performance) and reduced latency from 43 ms to 20ms (reducing 53% latency).
- Update model secondary face detection input size which increased analytic performance from 9 pipelines to 14 pipelines.
- We add a new feature to prepare for something big.
- We optimize TensorRT preprocess. In this part, we tried to implement it on our Vehicle Counting and reduce our latency from 29ms to 11ms for 1 pipeline, which reduce nearly 62% faster than before.
- HTTP request error handling in uncertainty dumper (for offline deployment).
- Mechanism to not send data for several tries to our cloud when failed once.
- Add support for fp16 inference.
- We implement it on our Vehicle Counting. We improve from 6 pipelines to 24 pipelines (increasing 300% performance) based on our T4 GPU and Intel Xeon 24 threads environment.
- Pipeline health check limit was increased to avoid errors when building our inference engine.
- We introduce a new mechanism on our Vehicle Counting to improve our performance. We improve our performance from 6 pipelines to 10 pipelines. Tested on our T4 and Intel Xeon 24 threads environment.
- We fix NMS and fix overlap not removing overlapped objects
Last modified 25d ago