Classification fixes (#20677)

* Don't run classification on stationary objects and set a maximum number of classifications

* Fix layout of classification selection

docs/docs/configuration/genai/review_summaries.md

@@ -39,6 +39,26 @@ Each installation and even camera can have different parameters for what is cons
 - Brief movement with legitimate items (bags, packages, tools, equipment) in appropriate zones is routine.
 ```
 
+### Image Source
+
+By default, review summaries use preview images (cached preview frames) which have a lower resolution but use fewer tokens per image. For better image quality and more detailed analysis, you can configure Frigate to extract frames directly from recordings at a higher resolution:
+
+```yaml
+review:
+  genai:
+    enabled: true
+    image_source: recordings # Options: "preview" (default) or "recordings"
+```
+
+When using `recordings`, frames are extracted at 480p resolution (480px height), providing better detail for the LLM while being mindful of context window size. This is particularly useful for scenarios where fine details matter, such as identifying license plates, reading text, or analyzing distant objects. Note that using recordings will:
+
+- Provide higher quality images to the LLM (480p vs 180p preview images)
+- Use more tokens per image (~200-300 tokens vs ~100 tokens for preview)
+- Result in fewer frames being sent to stay within context limits (typically 6-12 frames vs 8-20 frames)
+- Require that recordings are enabled for the camera
+
+If recordings are not available for a given time period, the system will automatically fall back to using preview frames.
+
 ### Additional Concerns
 
 Along with the concern of suspicious activity or immediate threat, you may have concerns such as animals in your garden or a gate being left open. These concerns can be configured so that the review summaries will make note of them if the activity requires additional review. For example:

docs/docs/configuration/reference.md

@@ -429,6 +429,10 @@ review:
     alerts: True
     # Optional: Enable GenAI review summaries for detections (default: shown below)
     detections: False
+    # Optional: Image source for GenAI (default: preview)
+    # Options: "preview" (uses cached preview frames at 180p) or "recordings" (extracts frames from recordings at 480p)
+    # Using "recordings" provides better image quality but uses ~2-3x more tokens per image (~200-300 vs ~100 tokens)
+    image_source: preview
     # Optional: Additional concerns that the GenAI should make note of (default: None)
     additional_concerns:
       - Animals in the garden

frigate/config/camera/review.py

@@ -1,10 +1,18 @@
+from enum import Enum
 from typing import Optional, Union
 
 from pydantic import Field, field_validator
 
 from ..base import FrigateBaseModel
 
-__all__ = ["ReviewConfig", "DetectionsConfig", "AlertsConfig"]
+__all__ = ["ReviewConfig", "DetectionsConfig", "AlertsConfig", "ImageSourceEnum"]
+
+
+class ImageSourceEnum(str, Enum):
+    """Image source options for GenAI Review."""
+
+    preview = "preview"
+    recordings = "recordings"
 
 
 DEFAULT_ALERT_OBJECTS = ["person", "car"]
@@ -77,6 +85,10 @@ class GenAIReviewConfig(FrigateBaseModel):
     )
     alerts: bool = Field(default=True, title="Enable GenAI for alerts.")
     detections: bool = Field(default=False, title="Enable GenAI for detections.")
+    image_source: ImageSourceEnum = Field(
+        default=ImageSourceEnum.preview,
+        title="Image source for review descriptions.",
+    )
     additional_concerns: list[str] = Field(
         default=[],
         title="Additional concerns that GenAI should make note of on this camera.",

frigate/data_processing/post/review_descriptions.py

@@ -3,6 +3,7 @@
 import copy
 import datetime
 import logging
+import math
 import os
 import shutil
 import threading
@@ -10,16 +11,18 @@ from pathlib import Path
 from typing import Any
 
 import cv2
+from peewee import DoesNotExist
 
 from frigate.comms.embeddings_updater import EmbeddingsRequestEnum
 from frigate.comms.inter_process import InterProcessRequestor
 from frigate.config import FrigateConfig
-from frigate.config.camera.review import GenAIReviewConfig
+from frigate.config.camera.review import GenAIReviewConfig, ImageSourceEnum
 from frigate.const import CACHE_DIR, CLIPS_DIR, UPDATE_REVIEW_DESCRIPTION
 from frigate.data_processing.types import PostProcessDataEnum
 from frigate.genai import GenAIClient
-from frigate.models import ReviewSegment
+from frigate.models import Recordings, ReviewSegment
 from frigate.util.builtin import EventsPerSecond, InferenceSpeed
+from frigate.util.image import get_image_from_recording
 
 from ..post.api import PostProcessorApi
 from ..types import DataProcessorMetrics
@@ -43,20 +46,35 @@ class ReviewDescriptionProcessor(PostProcessorApi):
         self.review_descs_dps = EventsPerSecond()
         self.review_descs_dps.start()
 
-    def calculate_frame_count(self) -> int:
-        """Calculate optimal number of frames based on context size."""
-        # With our preview images (height of 180px) each image should be ~100 tokens per image
-        # We want to be conservative to not have too long of query times with too many images
+    def calculate_frame_count(
+        self, image_source: ImageSourceEnum = ImageSourceEnum.preview
+    ) -> int:
+        """Calculate optimal number of frames based on context size and image source."""
         context_size = self.genai_client.get_context_size()
 
-        if context_size > 10000:
-            return 20
-        elif context_size > 6000:
-            return 16
-        elif context_size > 4000:
-            return 12
+        if image_source == ImageSourceEnum.recordings:
+            # With recordings at 480p resolution (480px height), each image uses ~200-300 tokens
+            # This is ~2-3x more than preview images, so we reduce frame count accordingly
+            # to avoid exceeding context limits and maintain reasonable inference times
+            if context_size > 10000:
+                return 12
+            elif context_size > 6000:
+                return 10
+            elif context_size > 4000:
+                return 8
+            else:
+                return 6
         else:
-            return 8
+            # With preview images (180px height), each image uses ~100 tokens
+            # We can send more frames since they're lower resolution
+            if context_size > 10000:
+                return 20
+            elif context_size > 6000:
+                return 16
+            elif context_size > 4000:
+                return 12
+            else:
+                return 8
 
     def process_data(self, data, data_type):
         self.metrics.review_desc_dps.value = self.review_descs_dps.eps()
@@ -88,36 +106,50 @@ class ReviewDescriptionProcessor(PostProcessorApi):
             ):
                 return
 
-            frames = self.get_cache_frames(
-                camera, final_data["start_time"], final_data["end_time"]
-            )
+            image_source = camera_config.review.genai.image_source
 
-            if not frames:
-                frames = [final_data["thumb_path"]]
-
-            thumbs = []
-
-            for idx, thumb_path in enumerate(frames):
-                thumb_data = cv2.imread(thumb_path)
-                ret, jpg = cv2.imencode(
-                    ".jpg", thumb_data, [int(cv2.IMWRITE_JPEG_QUALITY), 100]
+            if image_source == ImageSourceEnum.recordings:
+                thumbs = self.get_recording_frames(
+                    camera,
+                    final_data["start_time"],
+                    final_data["end_time"],
+                    height=480,  # Use 480p for good balance between quality and token usage
                 )
 
-                if ret:
-                    thumbs.append(jpg.tobytes())
-
-                if camera_config.review.genai.debug_save_thumbnails:
-                    id = data["after"]["id"]
-                    Path(os.path.join(CLIPS_DIR, "genai-requests", f"{id}")).mkdir(
+                if not thumbs:
+                    # Fallback to preview frames if no recordings available
+                    logger.warning(
+                        f"No recording frames found for {camera}, falling back to preview frames"
+                    )
+                    thumbs = self.get_preview_frames_as_bytes(
+                        camera,
+                        final_data["start_time"],
+                        final_data["end_time"],
+                        final_data["thumb_path"],
+                        id,
+                        camera_config.review.genai.debug_save_thumbnails,
+                    )
+                elif camera_config.review.genai.debug_save_thumbnails:
+                    # Save debug thumbnails for recordings
+                    Path(os.path.join(CLIPS_DIR, "genai-requests", id)).mkdir(
                         parents=True, exist_ok=True
                     )
-                    shutil.copy(
-                        thumb_path,
-                        os.path.join(
-                            CLIPS_DIR,
-                            f"genai-requests/{id}/{idx}.webp",
-                        ),
-                    )
+                    for idx, frame_bytes in enumerate(thumbs):
+                        with open(
+                            os.path.join(CLIPS_DIR, f"genai-requests/{id}/{idx}.jpg"),
+                            "wb",
+                        ) as f:
+                            f.write(frame_bytes)
+            else:
+                # Use preview frames
+                thumbs = self.get_preview_frames_as_bytes(
+                    camera,
+                    final_data["start_time"],
+                    final_data["end_time"],
+                    final_data["thumb_path"],
+                    id,
+                    camera_config.review.genai.debug_save_thumbnails,
+                )
 
             # kickoff analysis
             self.review_descs_dps.update()
@@ -231,6 +263,122 @@ class ReviewDescriptionProcessor(PostProcessorApi):
 
         return selected_frames
 
+    def get_recording_frames(
+        self,
+        camera: str,
+        start_time: float,
+        end_time: float,
+        height: int = 480,
+    ) -> list[bytes]:
+        """Get frames from recordings at specified timestamps."""
+        duration = end_time - start_time
+        desired_frame_count = self.calculate_frame_count(ImageSourceEnum.recordings)
+
+        # Calculate evenly spaced timestamps throughout the duration
+        if desired_frame_count == 1:
+            timestamps = [start_time + duration / 2]
+        else:
+            step = duration / (desired_frame_count - 1)
+            timestamps = [start_time + (i * step) for i in range(desired_frame_count)]
+
+        def extract_frame_from_recording(ts: float) -> bytes | None:
+            """Extract a single frame from recording at given timestamp."""
+            try:
+                recording = (
+                    Recordings.select(
+                        Recordings.path,
+                        Recordings.start_time,
+                    )
+                    .where((ts >= Recordings.start_time) & (ts <= Recordings.end_time))
+                    .where(Recordings.camera == camera)
+                    .order_by(Recordings.start_time.desc())
+                    .limit(1)
+                    .get()
+                )
+
+                time_in_segment = ts - recording.start_time
+                return get_image_from_recording(
+                    self.config.ffmpeg,
+                    recording.path,
+                    time_in_segment,
+                    "mjpeg",
+                    height=height,
+                )
+            except DoesNotExist:
+                return None
+
+        frames = []
+
+        for timestamp in timestamps:
+            try:
+                # Try to extract frame at exact timestamp
+                image_data = extract_frame_from_recording(timestamp)
+
+                if not image_data:
+                    # Try with rounded timestamp as fallback
+                    rounded_timestamp = math.ceil(timestamp)
+                    image_data = extract_frame_from_recording(rounded_timestamp)
+
+                if image_data:
+                    frames.append(image_data)
+                else:
+                    logger.warning(
+                        f"No recording found for {camera} at timestamp {timestamp}"
+                    )
+            except Exception as e:
+                logger.error(
+                    f"Error extracting frame from recording for {camera} at {timestamp}: {e}"
+                )
+                continue
+
+        return frames
+
+    def get_preview_frames_as_bytes(
+        self,
+        camera: str,
+        start_time: float,
+        end_time: float,
+        thumb_path_fallback: str,
+        review_id: str,
+        save_debug: bool,
+    ) -> list[bytes]:
+        """Get preview frames and convert them to JPEG bytes.
+
+        Args:
+            camera: Camera name
+            start_time: Start timestamp
+            end_time: End timestamp
+            thumb_path_fallback: Fallback thumbnail path if no preview frames found
+            review_id: Review item ID for debug saving
+            save_debug: Whether to save debug thumbnails
+
+        Returns:
+            List of JPEG image bytes
+        """
+        frame_paths = self.get_cache_frames(camera, start_time, end_time)
+        if not frame_paths:
+            frame_paths = [thumb_path_fallback]
+
+        thumbs = []
+        for idx, thumb_path in enumerate(frame_paths):
+            thumb_data = cv2.imread(thumb_path)
+            ret, jpg = cv2.imencode(
+                ".jpg", thumb_data, [int(cv2.IMWRITE_JPEG_QUALITY), 100]
+            )
+            if ret:
+                thumbs.append(jpg.tobytes())
+
+            if save_debug:
+                Path(os.path.join(CLIPS_DIR, "genai-requests", review_id)).mkdir(
+                    parents=True, exist_ok=True
+                )
+                shutil.copy(
+                    thumb_path,
+                    os.path.join(CLIPS_DIR, f"genai-requests/{review_id}/{idx}.webp"),
+                )
+
+        return thumbs
+
 
 @staticmethod
 def run_analysis(

frigate/data_processing/real_time/custom_classification.py

@@ -34,6 +34,8 @@ except ModuleNotFoundError:
 
 logger = logging.getLogger(__name__)
 
+MAX_OBJECT_CLASSIFICATIONS = 16
+
 
 class CustomStateClassificationProcessor(RealTimeProcessorApi):
     def __init__(
@@ -396,6 +398,18 @@ class CustomObjectClassificationProcessor(RealTimeProcessorApi):
         if obj_data.get("end_time") is not None:
             return
 
+        if obj_data.get("stationary"):
+            return
+
+        object_id = obj_data["id"]
+
+        if (
+            object_id in self.classification_history
+            and len(self.classification_history[object_id])
+            >= MAX_OBJECT_CLASSIFICATIONS
+        ):
+            return
+
         now = datetime.datetime.now().timestamp()
         x, y, x2, y2 = calculate_region(
             frame.shape,
@@ -427,7 +441,7 @@ class CustomObjectClassificationProcessor(RealTimeProcessorApi):
             write_classification_attempt(
                 self.train_dir,
                 cv2.cvtColor(crop, cv2.COLOR_RGB2BGR),
-                obj_data["id"],
+                object_id,
                 now,
                 "unknown",
                 0.0,
@@ -448,7 +462,7 @@ class CustomObjectClassificationProcessor(RealTimeProcessorApi):
         write_classification_attempt(
             self.train_dir,
             cv2.cvtColor(crop, cv2.COLOR_RGB2BGR),
-            obj_data["id"],
+            object_id,
             now,
             self.labelmap[best_id],
             score,
@@ -461,7 +475,7 @@ class CustomObjectClassificationProcessor(RealTimeProcessorApi):
         sub_label = self.labelmap[best_id]
 
         consensus_label, consensus_score = self.get_weighted_score(
-            obj_data["id"], sub_label, score, now
+            object_id, sub_label, score, now
         )
 
         if consensus_label is not None:
@@ -470,7 +484,7 @@ class CustomObjectClassificationProcessor(RealTimeProcessorApi):
                 == ObjectClassificationType.sub_label
             ):
                 self.sub_label_publisher.publish(
-                    (obj_data["id"], consensus_label, consensus_score),
+                    (object_id, consensus_label, consensus_score),
                     EventMetadataTypeEnum.sub_label,
                 )
             elif (
@@ -479,7 +493,7 @@ class CustomObjectClassificationProcessor(RealTimeProcessorApi):
             ):
                 self.sub_label_publisher.publish(
                     (
-                        obj_data["id"],
+                        object_id,
                         self.model_config.name,
                         consensus_label,
                         consensus_score,

web/src/components/overlay/detail/ObjectLifecycle.tsx

@@ -41,7 +41,7 @@ import {
   ContextMenuItem,
   ContextMenuTrigger,
 } from "@/components/ui/context-menu";
-import { useNavigate } from "react-router-dom";
+import { Link, useNavigate } from "react-router-dom";
 import { ObjectPath } from "./ObjectPath";
 import { getLifecycleItemDescription } from "@/utils/lifecycleUtil";
 import { IoPlayCircleOutline } from "react-icons/io5";
@@ -289,10 +289,10 @@ export default function ObjectLifecycle({
         timezone: config.ui.timezone,
         date_format:
           config.ui.time_format == "24hour"
-            ? t("time.formattedTimestampHourMinuteSecond.24hour", {
+            ? t("time.formattedTimestamp.24hour", {
                 ns: "common",
               })
-            : t("time.formattedTimestampHourMinuteSecond.12hour", {
+            : t("time.formattedTimestamp.12hour", {
                 ns: "common",
               }),
         time_style: "medium",
@@ -305,10 +305,10 @@ export default function ObjectLifecycle({
         timezone: config.ui.timezone,
         date_format:
           config.ui.time_format == "24hour"
-            ? t("time.formattedTimestampHourMinuteSecond.24hour", {
+            ? t("time.formattedTimestamp.24hour", {
                 ns: "common",
               })
-            : t("time.formattedTimestampHourMinuteSecond.12hour", {
+            : t("time.formattedTimestamp.12hour", {
                 ns: "common",
               }),
         time_style: "medium",
@@ -412,6 +412,7 @@ export default function ObjectLifecycle({
 
   return (
     <div className={className}>
+      <span tabIndex={0} className="sr-only" />
       {!fullscreen && (
         <div className={cn("flex items-center gap-2")}>
           <Button
@@ -649,10 +650,15 @@ export default function ObjectLifecycle({
                 </span>
                 {event.data?.recognized_license_plate && (
                   <>
-                    ·{" "}
-                    <span className="text-sm text-secondary-foreground">
-                      {event.data.recognized_license_plate}
-                    </span>
+                    <span className="text-secondary-foreground">·</span>
+                    <div className="text-sm text-secondary-foreground">
+                      <Link
+                        to={`/explore?recognized_license_plate=${event.data.recognized_license_plate}`}
+                        className="text-sm"
+                      >
+                        {event.data.recognized_license_plate}
+                      </Link>
+                    </div>
                   </>
                 )}
               </div>
@@ -832,10 +838,12 @@ function LifecycleIconRow({
           />
         </div>
 
-        <div className="flex w-full flex-row justify-between">
+        <div className="ml-2 flex w-full min-w-0 flex-1">
           <div className="flex flex-col">
-            <div>{getLifecycleItemDescription(item)}</div>
-            <div className="mt-1 flex flex-wrap items-center gap-2 text-sm text-secondary-foreground md:gap-5">
+            <div className="text-md flex items-start break-words text-left">
+              {getLifecycleItemDescription(item)}
+            </div>
+            <div className="mt-1 flex flex-wrap items-center gap-2 text-xs text-secondary-foreground md:gap-5">
               <div className="flex items-center gap-1">
                 <span className="text-primary-variant">
                   {t("objectLifecycle.lifecycleItemDesc.header.ratio")}
@@ -893,8 +901,9 @@ function LifecycleIconRow({
               )}
             </div>
           </div>
-
-          <div className={cn("p-1 text-sm")}>{formattedEventTimestamp}</div>
+        </div>
+        <div className="ml-3 flex-shrink-0 px-1 text-right text-xs text-primary-variant">
+          <div className="whitespace-nowrap">{formattedEventTimestamp}</div>
         </div>
       </div>
     </div>

web/src/components/timeline/DetailStream.tsx

@@ -22,6 +22,7 @@ import EventMenu from "@/components/timeline/EventMenu";
 import { FrigatePlusDialog } from "@/components/overlay/dialog/FrigatePlusDialog";
 import { cn } from "@/lib/utils";
 import { Tooltip, TooltipContent, TooltipTrigger } from "../ui/tooltip";
+import { Link } from "react-router-dom";
 
 type DetailStreamProps = {
   reviewItems?: ReviewSegment[];
@@ -499,15 +500,22 @@ function EventList({
               }}
               role="button"
             >
-              <span className="capitalize">{label}</span>
-              {event.data?.recognized_license_plate && (
-                <>
-                  ·{" "}
-                  <span className="text-sm text-secondary-foreground">
-                    {event.data.recognized_license_plate}
-                  </span>
-                </>
-              )}
+              <div className="flex gap-2">
+                <span className="capitalize">{label}</span>
+                {event.data?.recognized_license_plate && (
+                  <>
+                    <span className="text-secondary-foreground">·</span>
+                    <div className="text-sm text-secondary-foreground">
+                      <Link
+                        to={`/explore?recognized_license_plate=${event.data.recognized_license_plate}`}
+                        className="text-sm"
+                      >
+                        {event.data.recognized_license_plate}
+                      </Link>
+                    </div>
+                  </>
+                )}
+              </div>
             </div>
           </div>
           <div className="mr-2 flex flex-row justify-end">
@@ -615,10 +623,11 @@ function LifecycleItem({
           )}
         />
       </div>
-      <div className="flex w-full flex-row justify-between">
+
+      <div className="ml-0.5 flex min-w-0 flex-1">
         <Tooltip>
           <TooltipTrigger>
-            <div className="flex items-start text-left">
+            <div className="flex items-start break-words text-left">
               {getLifecycleItemDescription(item)}
             </div>
           </TooltipTrigger>
@@ -638,7 +647,9 @@ function LifecycleItem({
                   </span>
                   {areaPx !== undefined && areaPct !== undefined ? (
                     <span className="font-medium text-foreground">
-                      {areaPx} {t("pixels", { ns: "common" })} · {areaPct}%
+                      {areaPx} {t("pixels", { ns: "common" })}{" "}
+                      <span className="text-secondary-foreground">·</span>{" "}
+                      {areaPct}%
                     </span>
                   ) : (
                     <span>N/A</span>
@@ -648,7 +659,10 @@ function LifecycleItem({
             </div>
           </TooltipContent>
         </Tooltip>
-        <div className={cn("p-1 text-xs")}>{formattedEventTimestamp}</div>
+      </div>
+
+      <div className="ml-3 flex-shrink-0 px-1 text-right text-xs text-primary-variant">
+        <div className="whitespace-nowrap">{formattedEventTimestamp}</div>
       </div>
     </div>
   );

web/src/context/detail-stream-context.tsx

@@ -58,6 +58,11 @@ export function DetailStreamProvider({
     setAnnotationOffset(cfgOffset);
   }, [config, camera]);
 
+  // Clear selected objects when exiting detail mode or changing cameras
+  useEffect(() => {
+    setSelectedObjectIds([]);
+  }, [isDetailMode, camera]);
+
   const value: DetailStreamContextType = {
     selectedObjectIds,
     currentTime,

web/src/views/classification/ModelSelectionView.tsx

@@ -11,7 +11,6 @@ import {
   FrigateConfig,
 } from "@/types/frigateConfig";
 import { useEffect, useMemo, useState } from "react";
-import { isMobile } from "react-device-detect";
 import { useTranslation } from "react-i18next";
 import { FaFolderPlus } from "react-icons/fa";
 import { MdModelTraining } from "react-icons/md";
@@ -131,7 +130,7 @@ export default function ModelSelectionView({
           </Button>
         </div>
       </div>
-      <div className="flex size-full gap-2 p-2">
+      <div className="grid auto-rows-max grid-cols-2 gap-2 overflow-y-auto p-2 md:grid-cols-4 lg:grid-cols-5 xl:grid-cols-6 2xl:grid-cols-8 3xl:grid-cols-10">
         {selectedClassificationConfigs.length === 0 ? (
           <NoModelsView
             onCreateModel={() => setNewModel(true)}
@@ -208,14 +207,13 @@ function ModelCard({ config, onClick }: ModelCardProps) {
     <div
       key={config.name}
       className={cn(
-        "relative size-60 cursor-pointer overflow-hidden rounded-lg",
+        "relative aspect-square w-full cursor-pointer overflow-hidden rounded-lg",
         "outline-transparent duration-500",
-        isMobile && "w-full",
       )}
       onClick={() => onClick()}
     >
       <img
-        className={cn("size-full", isMobile && "w-full")}
+        className="size-full"
         src={`${baseUrl}clips/${config.name}/dataset/${coverImage?.name}/${coverImage?.img}`}
       />
       <ImageShadowOverlay />