---
title: Focaler-SIoU Loss in Object Detection
url: https://www.emergentmind.com/topics/focaler-siou-loss
type: topic
---

# Focaler-SIoU Loss in Object Detection

Below is an expository summary of the main constructions and results of Qian–Tudor’s “Differential Structure and Flow Equations on Rough Path Space.”  We work throughout with a Banach space V and fix 2<p<3.  

1.  Space of weakly‐geometric p-rough paths  
------------------------------------------------  
(a)  Tensor algebra up to level 2.  Let  
    T²(V) := 1⊕V⊕(V⊗V),  
   and write for any (s,t) with 0≤s≤t≤T a formal “increment”  
    X_{s,t} = (1, X^1_{s,t}, X^2_{s,t}) ∈ T²(V).  
(b)  Chen’s identity (multiplicativity).  We require for all 0≤s≤u≤t≤T  
    X_{s,t}  = X_{s,u} ⊗ X_{u,t},  
   i.e.  
    X^1_{s,t} = X^1_{s,u} + X^1_{u,t},  
    X^2_{s,t} = X^2_{s,u} + X^1_{s,u}⊗X^1_{u,t} + X^2_{u,t}.  
(c)  p-variation.  We say X has finite p-variation if for i=1,2  
    sup_{D\=\{s=t_0<⋯<t_N=t\}} ∑_{j} ‖X^i_{t_{j},t_{j+1}}‖^{p/i}  < ∞.  
(d)  Weakly geometric p-rough paths.  Denote by C^p(V) the set of all such multiplica­tive functionals.  Its subset WG C^p(V) consists of those X which arise as uniform limits of the canonical level-2 lifts of bounded‐variation paths.  One shows WG C^p(V) is a complete metric space under  
    d_p(X,Y) := \max_{i=1,2} \Bigl( \sup_{(s,t)} ‖X^i_{s,t}-Y^i_{s,t}‖^{p/i}\Bigr)^{i/p}.  

In particular, if x:[0,T]→V is C¹ then its canonical lift is  
    X^1_{s,t}=x_t−x_s,  
    X^2_{s,t}=\!\int_s^t (x_u−x_s)\otimes dx_u,  
and one checks this satisfies the above with equality in multiplicativity.  

2.  Smooth rough paths and directional derivatives  
------------------------------------------------------------  
To motivate a tangent calculus one first perturbs a *smooth* lift.  Let x,y be two C¹(V)–paths and form for ε∈ℝ the two‐parameter family of lifts X(ε) by  
    X(ε)^1_{s,t} = (x_t−x_s)+ε(y_t−y_s),  
    X(ε)^2_{s,t}  
    =  
     \int_s^t(x_u−x_s)+ε(y_u−y_s)\;\otimes\;d[x_u+ε\,y_u]  
    =\;X^2_{s,t}+ε\bigl[\!\int_s^t (x_u−x_s)\otimes dy_u + \int_s^t(y_u−y_s)\otimes dx_u\bigr]  
    \quad\;\;+\;\;ε^2\!\int_s^t (y_u−y_s)\otimes dy_u.  
Taking the Gateaux derivative at ε=0 one finds the *directional derivative* of the lift X↦X(ε) is itself a pair  
    D_yX = \bigl(0,\; \int_s^t(x_u−x_s)\otimes dy_u +\int_s^t(y_u−y_s)\otimes dx_u \bigr).  
Crucially, when x,y are only p-variation (non‐smooth) one cannot define these two cross‐integrals by Riemann sums alone.  Qian–Tudor show how to encode them as *projections* of a single higher‐dimensional rough path  
    Z ∈ WG C^p(V⊕V)  
whose components recover X,x–y and fill in the missing “area” terms T¹,²(Z) and T²,¹(Z).  

3.  Tangent space at X and equivalence of variations  
--------------------------------------------------------------  
(a)  Curves in WG C^p(V).  Any sufficiently regular curve ε↦C(ε)∈WG C^p(V) with C(0)=X admits a derivative in the linear space of continuous T²(V)–valued functions.  One writes  
    C(ε)^1_{s,t}=X^1_{s,t}+ε A^1_{s,t}+o(ε),  
    C(ε)^2_{s,t}=X^2_{s,t}+ε A^2_{s,t}+o(ε).  
(b)  Abstract tangent vector.  One proves there is a *unique* equivalence class of pairs  
    [\,Z\,,\,p\,]\;,  
   with Z∈WG C^p(V⊕V), p∈WG C^{p/2}(V⊗V), and T¹(Z)=X, which reproduces the first‐level increment A¹= T²(Z)¹, and second‐level A²= T¹,²(Z)+T²,¹(Z)+p.  
(c)  Equivalence relation.  Two data (Z,p)∼(Ŷ,q) if the associated “variational curves”  
    ε↦V(Z,p)(ε)  and  ε↦V(Ŷ,q)(ε)  
have the same Gateaux derivative at ε=0 in the p-variation topology.  
(d)  Tangent space.  Denote by T_XWG C^p(V) the collection of all such classes [Z,p].  In particular one recovers the usual Cameron–Martin directions (for Brownian rough path) as a proper subspace, but the tangent space is *strictly larger* — every *weakly‐geometric* perturbation of X gives a tangent.  

4.  The tangent bundle  
-------------------------------  
Putting the fibers T_XWG C^p(V) together as X varies yields what we may call the *tangent bundle* TWG C^p(V)→WG C^p(V).  Though it is not a smooth finite‐dimensional bundle, it carries a natural metric  
    d_{T,p}\bigl([Z,p]\,,\,[W,q]\bigr)  
        := \max\Bigl\{d_p\bigl(T¹(Z),T¹(W)\bigr),\,\| T^{1,2}(Z)+T^{2,1}(Z)+p \;-\; T^{1,2}(W)-T^{2,1}(W)-q\|_{p/2}\Bigr\},  
and a Banach‐vector‐space structure on each fiber.  One checks that the classical Cameron–Martin space embeds isometrically in each fiber, but there are infinitely many further directions.  

5.  Flow equations on WG C^p(V)  
-----------------------------------------------  
Given a *vector‐field*  
    F: WG C^p(V) → TWG C^p(V), X↦[\,Z(X)\,,\,p(X)\,],  
one asks for a path U:[0,τ]→WG C^p(V) solving the *rough‐path flow*  
    \frac{d}{dt}U(t) = F\bigl(U(t)\bigr),   U(0)=X_0.  
Here “d/dt” is interpreted so that infinitesimally  
    U(t+ε) ≃ V\bigl(Z(U(t)),p(U(t))\bigr)(ε).  
(a)  Local and global Lipschitz.  One shows that if F is *locally Lipschitz* in the sense  
    d_p\bigl(Z(X),Z(Y)\bigr)+\|p(X)-p(Y)\|_{p/2}  ≤ C\,d_p(X,Y),  
(and a corresponding bound in some q‐variation norm with q>p)  
then there is a unique *local* solution on some [0,a].  If these Lipschitz constants are uniform in X, one obtains a *global* solution for all t≥0.  
(b)  Existence via Euler scheme.  One constructs an ε–approximate solution by “following the variational curve” over small subintervals [t_i,t_{i+1}] of size ≲ε, then letting ε→0 and passing to a limit in d_q.  Uniqueness follows by a Grönwall‐type estimate in the tangent‐space metric.  

6.  Examples and recovery of classical ODEs/SDEs  
---------------------------------------------------------  
•  If X is the canonical lift of a smooth signal x and one chooses F(X) induced by a usual vector‐field V:V→L(V,V) via the Lyons–Itô map, then U(t) coincides with the (lifted) solution of the ordinary ODE dx_t=V(x_t)dx.  
•  If X is the Brownian rough path (a.s. 2<p<3), then the flow reduces to the solution of a Stratonovich SDE dY=V(Y)○dW.  In this way the theory is *pathwise* and does not depend on probabilistic quasi‐invariance arguments.  

7.  Position in rough path analysis and applications  
--------------------------------------------------------------  
Qian–Tudor’s calculus on WG C^p(V) grants a genuine “tangent‐bundle” and “flow” theory directly on rough path space.  Unlike Malliavin calculus, no reference measure (e.g. Wiener) or Cameron–Martin quasi‐invariance is required.  This framework:  
•  Provides *deterministic* variational formulas for functionals of rough paths;  
•  Extends Malliavin‐type differentiation beyond the Cameron–Martin directions;  
•  Yields new existence/uniqueness results for ODEs driven by rough signals;  
•  Lends itself to pathwise sensitivity analysis of SDE solutions;  
•  Opens the door to geometric analysis (curvature, torsion) on rough path bundles.  

In summary, the paper introduces a *full differential structure* on the space of weakly‐geometric p-rough paths, identifies a canonically large tangent space at each X, and uses it to formulate and solve *flow equations* dU=F(U)dX entirely at the rough‐path level.

Source: https://www.emergentmind.com/topics/focaler-siou-loss