remove possibility of using other kf than xysr

boxmot/configs/deepocsort.yaml

@@ -13,5 +13,4 @@ inertia: 0.2
 iou_thresh: 0.3
 max_age: 30
 min_hits: 1
-new_kf_off: false
 w_association_emb: 0.75

boxmot/motion/kalman_filters/xysr_kf.py

@@ -108,42 +108,29 @@ def __init__(self, dim_x, dim_z, dim_u=0, max_obs=50):
         self.observed = False
 
 
-    def apply_affine_correction(self, m, t, new_kf):
+    def apply_affine_correction(self, m, t):
         """
         Apply to both last state and last observation for OOS smoothing.
 
         Messy due to internal logic for kalman filter being messy.
         """
-        if new_kf:
-            big_m = np.kron(np.eye(4, dtype=float), m)
-            self.x = big_m @ self.x
-            self.x[:2] += t
-            self.P = big_m @ self.P @ big_m.T
-
-            # If frozen, also need to update the frozen state for OOS
-            if not self.observed and self.attr_saved is not None:
-                self.attr_saved["x"] = big_m @ self.attr_saved["x"]
-                self.attr_saved["x"][:2] += t
-                self.attr_saved["P"] = big_m @ self.attr_saved["P"] @ big_m.T
-                self.attr_saved["last_measurement"][:2] = m @ self.attr_saved["last_measurement"][:2] + t
-                self.attr_saved["last_measurement"][2:] = m @ self.attr_saved["last_measurement"][2:]
-        else:
-            scale = np.linalg.norm(m[:, 0])
-            self.x[:2] = m @ self.x[:2] + t
-            self.x[4:6] = m @ self.x[4:6]
+
+        scale = np.linalg.norm(m[:, 0])
+        self.x[:2] = m @ self.x[:2] + t
+        self.x[4:6] = m @ self.x[4:6]
 
-            self.P[:2, :2] = m @ self.P[:2, :2] @ m.T
-            self.P[4:6, 4:6] = m @ self.P[4:6, 4:6] @ m.T
+        self.P[:2, :2] = m @ self.P[:2, :2] @ m.T
+        self.P[4:6, 4:6] = m @ self.P[4:6, 4:6] @ m.T
 
-            # If frozen, also need to update the frozen state for OOS
-            if not self.observed and self.attr_saved is not None:
-                self.attr_saved["x"][:2] = m @ self.attr_saved["x"][:2] + t
-                self.attr_saved["x"][4:6] = m @ self.attr_saved["x"][4:6]
+        # If frozen, also need to update the frozen state for OOS
+        if not self.observed and self.attr_saved is not None:
+            self.attr_saved["x"][:2] = m @ self.attr_saved["x"][:2] + t
+            self.attr_saved["x"][4:6] = m @ self.attr_saved["x"][4:6]
 
-                self.attr_saved["P"][:2, :2] = m @ self.attr_saved["P"][:2, :2] @ m.T
-                self.attr_saved["P"][4:6, 4:6] = m @ self.attr_saved["P"][4:6, 4:6] @ m.T
+            self.attr_saved["P"][:2, :2] = m @ self.attr_saved["P"][:2, :2] @ m.T
+            self.attr_saved["P"][4:6, 4:6] = m @ self.attr_saved["P"][4:6, 4:6] @ m.T
 
-                self.attr_saved["last_measurement"][:2] = m @ self.attr_saved["last_measurement"][:2] + t
+            self.attr_saved["last_measurement"][:2] = m @ self.attr_saved["last_measurement"][:2] + t
 
 
     def predict(self, u=None, B=None, F=None, Q=None):
@@ -334,33 +321,6 @@ def likelihood(self):
 
         return self._likelihood
 
-    def rts_smoother(self, Xs, Ps, Fs=None, Qs=None):
-        """ Runs the Rauch-Tung-Striebel (RTS) smoother on a set of
-        Kalman filter output, consisting of means and covariances.
-        """
-
-        if len(Xs) != len(Ps):
-            raise ValueError("length of Xs and Ps must be equal")
-
-        n = len(Xs)
-        dim_x = Xs[0].shape[0]
-
-        if Fs is None:
-            Fs = [self.F] * n
-        if Qs is None:
-            Qs = [self.Q] * n
-
-        # smoother gain
-        K = np.zeros((n, dim_x, dim_x))
-
-        x, P = Xs.copy(), Ps.copy()
-        for k in range(n-2, -1, -1):
-            P_pred = dot(dot(Fs[k], P[k]), Fs[k].T) + Qs[k]
-            K[k] = dot(dot(P[k], Fs[k].T), self.inv(P_pred))
-            x[k] += dot(K[k], x[k+1] - dot(Fs[k], x[k]))
-            P[k] += dot(dot(K[k], P[k+1] - P_pred), K[k].T)
-        return (x, P, K)
-
     def batch_filter(self, zs, us=None, Bs=None, Fs=None, Qs=None, Hs=None, Rs=None):
         """ Batch processes a sequence of measurements.
 

boxmot/trackers/deepocsort/deep_ocsort.py

@@ -73,35 +73,20 @@ def speed_direction(bbox1, bbox2):
     return speed / norm
 
 
-def new_kf_process_noise(w, h, p=1 / 20, v=1 / 160):
-    Q = np.diag(
-        ((p * w) ** 2, (p * h) ** 2, (p * w) ** 2, (p * h) ** 2, (v * w) ** 2, (v * h) ** 2, (v * w) ** 2, (v * h) ** 2)
-    )
-    return Q
-
-
-def new_kf_measurement_noise(w, h, m=1 / 20):
-    w_var = (m * w) ** 2
-    h_var = (m * h) ** 2
-    R = np.diag((w_var, h_var, w_var, h_var))
-    return R
-
-
 class KalmanBoxTracker(object):
     """
     This class represents the internal state of individual tracked objects observed as bbox.
     """
 
     count = 0
 
-    def __init__(self, det, delta_t=3, emb=None, alpha=0, new_kf=False, max_obs=50):
+    def __init__(self, det, delta_t=3, emb=None, alpha=0, max_obs=50):
         """
         Initialises a tracker using initial bounding box.
 
         """
         # define constant velocity model
         self.max_obs=max_obs
-        self.new_kf = new_kf
         bbox = det[0:5]
         self.conf = det[4]
         self.cls = det[5]
@@ -200,11 +185,8 @@ def update(self, det):
             self.time_since_update = 0
             self.hits += 1
             self.hit_streak += 1
-            if self.new_kf:
-                R = new_kf_measurement_noise(self.kf.x[2, 0], self.kf.x[3, 0])
-                self.kf.update(self.bbox_to_z_func(bbox), R=R)
-            else:
-                self.kf.update(self.bbox_to_z_func(bbox))
+
+            self.kf.update(self.bbox_to_z_func(bbox))
         else:
             self.kf.update(det)
             self.frozen = True
@@ -233,27 +215,16 @@ def apply_affine_correction(self, affine):
                 self.observations[self.age - dt][:4] = ps.T.reshape(-1)
 
         # Also need to change kf state, but might be frozen
-        self.kf.apply_affine_correction(m, t, self.new_kf)
+        self.kf.apply_affine_correction(m, t)
 
     def predict(self):
         """
         Advances the state vector and returns the predicted bounding box estimate.
         """
         # Don't allow negative bounding boxes
-        if self.new_kf:
-            if self.kf.x[2] + self.kf.x[6] <= 0:
-                self.kf.x[6] = 0
-            if self.kf.x[3] + self.kf.x[7] <= 0:
-                self.kf.x[7] = 0
-
-            # Stop velocity, will update in kf during OOS
-            if self.frozen:
-                self.kf.x[6] = self.kf.x[7] = 0
-            Q = new_kf_process_noise(self.kf.x[2, 0], self.kf.x[3, 0])
-        else:
-            if (self.kf.x[6] + self.kf.x[2]) <= 0:
-                self.kf.x[6] *= 0.0
-            Q = None
+        if (self.kf.x[6] + self.kf.x[2]) <= 0:
+            self.kf.x[6] *= 0.0
+        Q = None
 
         self.kf.predict(Q=Q)
         self.age += 1
@@ -294,7 +265,6 @@ def __init__(
         embedding_off=False,
         cmc_off=False,
         aw_off=False,
-        new_kf_off=False,
         **kwargs
     ):
         super().__init__(max_age=max_age)
@@ -323,7 +293,6 @@ def __init__(
         self.embedding_off = embedding_off
         self.cmc_off = cmc_off
         self.aw_off = aw_off
-        self.new_kf_off = new_kf_off
 
     @PerClassDecorator
     def update(self, dets: np.ndarray, img: np.ndarray, embs: np.ndarray = None) -> np.ndarray:
@@ -469,7 +438,6 @@ def update(self, dets: np.ndarray, img: np.ndarray, embs: np.ndarray = None) ->
                 delta_t=self.delta_t,
                 emb=dets_embs[i],
                 alpha=dets_alpha[i],
-                new_kf= False,
                 max_obs=self.max_obs
             )
             self.active_tracks.append(trk)

tracking/evolve.py

@@ -183,7 +183,6 @@ def get_new_config(self, trial):
             embedding_off = trial.suggest_categorical("embedding_off", [True, False])
             cmc_off = trial.suggest_categorical("cmc_off", [True, False])
             aw_off = trial.suggest_categorical("aw_off", [True, False])
-            new_kf_off = trial.suggest_categorical("new_kf_off", [True, False])
 
             d = {
                 'det_thresh': det_thresh,
@@ -199,7 +198,6 @@ def get_new_config(self, trial):
                 'embedding_off': embedding_off,
                 'cmc_off': cmc_off,
                 'aw_off': aw_off,
-                'new_kf_off': new_kf_off
             }
 
         # overwrite existing config for tracker