zishuo/frigate
1
0
Fork 0
mirror of https://github.com/blakeblackshear/frigate.git synced 2025-09-26 19:41:29 +08:00
Code Issues Packages Projects Releases Wiki Activity

Improve 640x640 model detection of small objects (#20190)

Browse Source 
* Allow larger models to have smaller regions

* remove unnecessary hailo resize

* Update benchmark

* Fix table

* Update nvidia specs
This commit is contained in:
Nicolas Mowen
2025-09-23 14:49:54 -06:00
committed by GitHub
parent 2f99a17e64
commit 7e2f5a3017
3 changed files with 26 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
docs/docs/frigate/hardware.md
View File

@@ -180,10 +180,10 @@ Inference speeds will vary greatly depending on the GPU and the model used.
✅ - Accelerated with CUDA Graphs
❌ - Not accelerated with CUDA Graphs
| Name | ✅ YOLOv9 Inference Time | ✅ RF-DETR Inference Time | ❌ YOLO-NAS Inference Time
| --------------- | ------------------------ | ------------------------- | -------------------------- |
| RTX 3050 | t-320: 8 ms s-320: 10 ms | Nano-320: ~ 12 ms | 320: ~ 10 ms 640: ~ 16 ms |
| RTX 3070 | t-320: 6 ms s-320: 8 ms | Nano-320: ~ 9 ms | 320: ~ 8 ms 640: ~ 14 ms |
| Name | ✅ YOLOv9 Inference Time | ✅ RF-DETR Inference Time | ❌ YOLO-NAS Inference Time |
| --------------- | ------------------------------------- | ------------------------- | --------------------------- |
| RTX 3050 | t-320: 8 ms s-320: 10 ms s-640: 28 ms | Nano-320: ~ 12 ms | 320: ~ 10 ms 640: ~ 16 ms |
| RTX 3070 | t-320: 6 ms s-320: 8 ms s-640: 25 ms | Nano-320: ~ 9 ms | 320: ~ 8 ms 640: ~ 14 ms |
| RTX A4000 | | | 320: ~ 15 ms |
| Tesla P40 | | | 320: ~ 105 ms |
@@ -198,8 +198,9 @@ Apple Silicon can not run within a container, so a ZMQ proxy is utilized to comm
:::
| Name | YOLOv9 Inference Time |
| --------- | ---------------------- |
| M3 Pro | t-320: 6 ms s-320: 8ms |
| --------- | ------------------------------------ |
| M4 | s-20: 10 ms |
| M3 Pro | t-320: 6 ms s-320: 8 ms s-640: 20 ms |
| M1 | s-320: 9ms |
### ROCm - AMD GPU

4
frigate/detectors/plugins/hailo8l.py
View File

@@ -33,10 +33,6 @@ def preprocess_tensor(image: np.ndarray, model_w: int, model_h: int) -> np.ndarr
image = image[0]
h, w = image.shape[:2]
if (w, h) == (320, 320) and (model_w, model_h) == (640, 640):
return cv2.resize(image, (model_w, model_h), interpolation=cv2.INTER_LINEAR)
scale = min(model_w / w, model_h / h)
new_w, new_h = int(w * scale), int(h * scale)
resized_image = cv2.resize(image, (new_w, new_h), interpolation=cv2.INTER_CUBIC)

15
frigate/util/object.py
View File

@@ -269,7 +269,20 @@ def is_object_filtered(obj, objects_to_track, object_filters):
def get_min_region_size(model_config: ModelConfig) -> int:
"""Get the min region size."""
return max(model_config.height, model_config.width)
largest_dimension = max(model_config.height, model_config.width)
if largest_dimension > 320:
# We originally tested allowing any model to have a region down to half of the model size
# but this led to many false positives. In this case we specifically target larger models
# which can benefit from a smaller region in some cases to detect smaller objects.
half = int(largest_dimension / 2)
if half % 4 == 0:
return half
return int((half + 3) / 4) * 4
return largest_dimension
def create_tensor_input(frame, model_config: ModelConfig, region):