Ck/ablation query

data/evaluation/ips.json

@@ -1,7 +1,7 @@
 [
   "34.72.31.210",
   "34.31.173.29",
-  "34.70.115.134",
+  "35.239.188.147",
   "34.28.144.66",
   "34.173.162.143",
   "35.202.139.99",

optimized_ingestion/stages/detection_estimation/__init__.py

@@ -85,10 +85,10 @@ def _run(self, payload: "Payload"):
             next_frame_num = i + 1
 
             det, _, dids = dets[i]
-            # if objects_count_change(dets, i, i + 5) <= i + 1:
-            #     # will not map segment if cannot skip in the first place
-            #     metadata.append([])
-            #     continue
+            if new_car(dets, i, i + 5) <= i + 1:
+                # will not map segment if cannot skip in the first place
+                metadata.append([])
+                continue
 
             start_detection_time = time.time()
             logger.info(f"current frame num {i}")
@@ -110,7 +110,7 @@ def _run(self, payload: "Payload"):
                                                             fps=current_fps)
             total_sample_plan_time.append(time.time() - start_generate_sample_plan)
             next_frame_num = next_sample_plan.get_next_frame_num()
-            # next_frame_num = objects_count_change(dets, i, next_frame_num)
+            next_frame_num = new_car(dets, i, next_frame_num)
             logger.info(f"founded next_frame_num {next_frame_num}")
             metadata.append(all_detection_info)
 
@@ -121,7 +121,7 @@ def _run(self, payload: "Payload"):
 
         # TODO: ignore the last frame ->
         metadata.append([])
-        skipped_frame_num.append(len(payload.video) - 1)
+        # skipped_frame_num.append(len(payload.video) - 1)
 
         #     times.append([t2 - t1 for t1, t2 in zip(t[:-1], t[1:])])
         # logger.info(np.array(times).sum(axis=0))
@@ -148,6 +148,15 @@ def _run(self, payload: "Payload"):
         return keep, {DetectionEstimation.classname(): metadata}
 
 
+def new_car(dets: "list[D2DMetadatum]", cur: "int", nxt: "int"):
+    len_det = len(dets[cur][0])
+    for j in range(cur + 1, min(nxt + 1, len(dets))):
+        future_det = dets[j][0]
+        if len(future_det) > len_det:
+            return j
+    return nxt
+
+
 def objects_count_change(dets: "list[D2DMetadatum]", cur: "int", nxt: "int"):
     det, _, _ = dets[cur]
     for j in range(cur + 1, min(nxt + 1, len(dets))):

optimized_ingestion/stages/detection_estimation/detection_estimation.py

@@ -42,7 +42,8 @@
     trajectory_3d,
 )
 
-MAX_SKIP = 5
+# MAX_SKIP = 5
+MAX_SKIP = 1000
 
 
 @dataclass
@@ -160,7 +161,7 @@ class SamplePlan:
     action: "Action | None" = None
 
     def update_next_sample_frame_num(self):
-        next_sample_frame_num = self.next_frame_num + MAX_SKIP - 1
+        next_sample_frame_num = min(self.next_frame_num + MAX_SKIP - 1, len(self.video))
         assert self.all_detection_info is not None
         for detection_info in self.all_detection_info:
             # get the frame num of car exits,

playground/run-ablation.ipynb

@@ -357,11 +357,11 @@
     "    }\n",
     "    print('# of total    videos:', len(videos))\n",
     "\n",
-    "    names = set(sampled_scenes[:150])\n",
+    "    names = set(sampled_scenes[:80])\n",
     "    # names = set(sampled_scenes)\n",
     "    filtered_videos = [\n",
     "        n for n in videos\n",
-    "        if n[6:10] in names and n.endswith('FRONT')\n",
+    "        if n[6:10] in names and 'FRONT' in n # and n.endswith('FRONT')\n",
     "    ]\n",
     "    N = len(filtered_videos)\n",
     "    print('# of filtered videos:', N)\n",