# QIMP SDK release 1.4

## Release information

Table : Software version

| Software | Version |
| --- | --- |
| Yocto | Scarthgap 5.0.6 |
| Kernel | 6.6.65 |

Table : Release tag version

| Release tag | Version |
| --- | --- |
| Firmware release tag | r1.0\_00075.0 |
| Manifest release tag | qcom-6.6.65-QLI.1.4-Ver.1.1.xml |
| Meta-qcom-extras release tag | r1.0\_00077.0 |
| QIMP SDK release tag | qcom-6.6.65-QLI.1.4-Ver.1.1\_qim-product-sdk-1.1.2.xml |

Table : Supported platforms and reference kits

| SoC platforms | Reference kits |
| --- | --- |
| QCS6490 | <ul class="simple"><br><li><p>Qualcomm<sup>®</sup> RB3 Gen 2 Vision Development Kit</p></li><br><li><p>Qualcomm<sup>®</sup> RB3 Gen 2 Core Development Kit</p></li><br></ul> |
| QCS5430 | <ul class="simple"><br><li><p>Qualcomm<sup>®</sup> RB3 Gen 2 Lite Vision Development Kit</p></li><br><li><p>Qualcomm<sup>®</sup> RB3 Gen 2 Lite Core Development Kit</p></li><br></ul> |
| QCS9075 | Qualcomm Dragonwing^TM^ IQ-9075 Evaluation Kit (EVK) |
| QCS8275 | Qualcomm^®^ IQ8 Beta EVK |

## Contents of the release

The contents of the Qualcomm^®^ Intelligent Multimedia Product (QIMP) SDK release include:

- Recipes for building the individual components:

    - Qualcomm^®^ Intelligent Multimedia SDK (IM SDK)
    - Lite Runtime (LiteRT)

Note

LiteRT was formerly known as TensorFlow Lite.
    - Qualcomm^®^ Neural Processing SDK
    - Qualcomm^®^ AI Engine direct SDK
- Sample applications that demonstrate how to use the Qualcomm IM SDK to develop AI edge-based applications.

To get started with the QIMP SDK, see [Qualcomm Intelligent Multimedia Product (QIMP) SDK Quick Start Guide](https://docs.qualcomm.com/bundle/publicresource/topics/80-70018-51/introduction.html).

## New features

The following new features applicable to QCS6490, QCS9075, and QCS8275 are introduced in the QIMP SDK release:

- Microservices:

    - Yolo V8-based person detection.
    - Face detection or recognition for access control.
    - General purpose daisy chain.
- AI-driven safety and monitoring system.

    - **Restricted zone alerts**: Define zones in the camera’s FOV to generate alerts when someone enters these areas.
    - **PPE compliance**: Ensure safety gear compliance with alerts for missing or restricted items, configurable using a web interface.
- As part of the Yocto upgrade to Scarthgap, the GStreamer version has been upgraded to 1.22.
- Qualcomm plugins utilize a new color format, NV12\_Q08C, which is introduced for NV12 with universal band width compression (UBWC). This color format replaces the previous method of using flags such as compression=ubwc.
- V4L2 supports `dmabuf` and `dmabuf-import` modes for efficient zero-copy data transfer.
- Audio AI: End-to-end support for general audio classification including sampling, preprocessing, inference, and post-processing.
- The docker base image is updated to Ubuntu 24.04 and the GStreamer version within the docker environment is 1.24.

**For QCS9075:**

- AI/ML parallel inference supports up to 32 streams.

## Sample applications

| Sample application | Description | Supported SoC |
| --- | --- | --- |
| **AI/ML applications** | **AI/ML applications** | **AI/ML applications** |
| `gst-ai-audio-classification` | Classification on streams from audio source. | QCS6490, QCS9075, and QCS8275 |
| `gst-ai-metadata-parser-example` | Parses metadata using appsink plugin on streams from camera, file, or RTSP sources. Also, it provides the human count from the stream. | QCS6490, QCS9075, and QCS8275 |
| `gst-ai-usb-camera-app` | USB single camera streaming for preview, video encoder, or network (RTSP) along with object detection and preview. | QCS6490, QCS9075, and QCS8275 |
| **Video application** | **Video application** | **Video application** |
| `gst-opencv-transform` | Video playback using opencv API. Parses input video file, captures frame using cv video capture, and streams videoplay and composites to display using waylandsink. | QCS6490, QCS9075, and QCS8275 |
| `gst-jpg-image-decode` | Decodes JPEG images where user can view decoded images on waylandsink. | QCS6490, QCS9075, and QCS8275 |
| `gst-camera-opencv-resize` | Live camera stream using opencv API, which includes color conversion and resize. The application captures camera frame as an input using cv video capture, converts to RGBA using downstream plugin, and resizes to user–specified resolution. Resize output displays on screen. | QCS6490, QCS9075, and QCS8275 |
| `gst-jpg-decode-example` | Collects the live video input from a camera, file, or an RTSP stream and uses the Qualcomm^®^ Neural Network face detection model to produce a preview with the overlaid AI model output on the HDMI display. | QCS6490, QCS9075, and QCS8275 |

Note

`gst-usb-camera-example` is removed from multimedia sample applications. Use `gst-ai-usb-camera-app` instead of `gst-usb-camera-example`.

For the complete list of sample applications supported in the QIMP SDK and instructions on how to run them, see [Sample applications](https://docs.qualcomm.com/bundle/publicresource/topics/80-70018-50/example-applications.html).

## Issues resolved

The following issues are resolved in the QIMP SDK release:

- A drop of 1 to 2 fps may be observed with the three-stream camera use case.
- Concurrent streams with smart codec functionality encounter cleanup errors at end of stream.
- Pose estimation for the `facemap_3dmm_quantized.bin` (Qualcomm Neural Network) AI hub model doesn’t work with the Qualcomm IM SDK pipeline due to caps mismatch.
- Caps mismatch with the AI hub face detection Qualcomm Neural Network model.
- Support for the FastCV engine is disabled in qtivtransform.

## Limitations

The following are the known limitations in the QIMP SDK release:

- When using the `qtioverlay` plugin with detection models, frame drops may occur, especially with many detections.

    **Workaround**: Use `qtivcomposer` for detection-based ML use cases.
- AI/ML parallel inference for 32 streams varies between 20–30 fps depending on the model and input stream, which is less than the expected 30 fps.
- Segfault occurs while using Ctrl + C for Qualcomm^®^ Neural Processing use cases with DLC models.
- Frame drops are observed with the Qualcomm Neural Network plugin while running on GPU delegate.
- Low FPS is observed with GStreamer pipelines for daisychain and pose detection.
- Stability issues are observed with GStreamer pipeline in batched model use cases.
- High inference time is observed with `deeplabv3_resnet50.dlc`.
- Concurrent streams with smart codec functionality encounter pipeline stalls at the end of the stream.
- Reverse playback is limited to video streams with a GOP length that fits within the buffer limitations of the video driver (&lt;= 26).
- Support for GStreamer daemon (GStd), NV12 UBWC compressed streams, and sample apps dependent on NV12 UBWC aren’t enabled inside the docker.
- The `gst-ai-parallel-inference` sample application hangs for QCS8275.
- For QCS9075 and QCS8275:

    - A hang is observed while running use cases with batched models using GStreamer.
    - Random corruption is observed at the beginning of playback for NV12 UBWC compressed low-resolution streams (480p).

Last Published: Apr 09, 2025

[Next Topic
QIMP SDK release 1.3](https://docs.qualcomm.com/bundle/publicresource/80-70018-52/topics/qimp_sdk_release_1_3.md)

Source: [https://docs.qualcomm.com/doc/80-70018-52/topic/qimp_sdk_release_1_4.html](https://docs.qualcomm.com/doc/80-70018-52/topic/qimp_sdk_release_1_4.html)