| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
Method no longer takes `robustCheck` parameter. Instead,
FalsingManager#isSimpleTap is added for basic checking, and
FalsingManager#isTap does robust checking by default.
FalsingManager#isTap takes an enum value for penalty instead of
a double, making the value more understandable.
Bug: 172655679
Test: atest SystemUITests && manual
Change-Id: Ib4a99f87bcd6acee67a98420f460c98d44fa6360
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Makes the FalsingManager slightly less stateful (passing the
interaction type directly to the classifiers as needed). It also
includes more logging than we had before, listing all the failed
classifiers for a gesture, along with the HistoryTracker's
belief and confidence.
Logs are now delayed one gesture. Instead of immediately logging when
a falsing call is made, it waits until the gesture is marked as
complete. This confers two advantages:
1) If multiple calls to #isFalseGesture (and similar) are made, we
don't log multiple times. We only log the result once.
2) It allows us to log the effects on the HistoryTracker, as that only
updates itself when the gesture is marked as completed.
Bug: 172655679
Test: atest SystemUITests && manual
Change-Id: I653ac2abb03a91ff000b46075e58137b04226023
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
| |
Quicksettings buttons will now reject single taps that either don't
look like clean single taps or that fail other single-tap falsing
criteria.
However, with this change we also allow clean double taps to be
recognized as single taps for the purposes of falsing. If a single
tap fails, a double tap might work.
Bug: 172655679
Test: atest SystemUITests && manual
Change-Id: Icab57e98e2784f3c152b4496db8fc97c2235cbec
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This uses the HistoryTracker on pin-based input to determine if a
tap outside of an intended region is a false touch, retracting the
pin input if we believe pocket dialing is occuring.
To do this, the cl properly integrates the HistoryTracker into
falsing, at least for single taps.
Most importantly, HistoryTracker#falsingBelief now return 0.5 when
it is unsure if a tap is false or not, and tends towards 0 when it
believes it's valid and 1 when it believes it's false.
HistoryTracker#falsingConfidence remains unchanged.
Test: atest SystemUITests && manual
Bug: 172655679
Change-Id: Ie771b1bf8ac564af7ffb68e190772fff5c562e89
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
We were keeping an unbounded list of MotionEvents inside of the
FalsingManager. Generally, this list should have shrunk itself
periodically, but in practice it only shrank itself when someone
double taps a notification.
Meanwhile, we weren't even using the data in the list!
With this change, we only keep the current and prior list of
MotionEvents - the ones we actually use. The list of historical
events is removed, so the leak goes with it.
Fixes: 177329773
Test: atest SystemUITests && manual
Change-Id: If8bfeea944850d972434915018b5bbfb3acb80e4
|
| |
|
|
|
|
|
|
|
|
| |
Clasifiers now return a Result object on all their classification
methods. The Result object now contains the "reason" for falsing
instead of asking the Classifier for the reason after the fact.
Bug: 172655679
Test: atest SystemUITests
Change-Id: I0865f18cbae9367c203936fbbc3923de55c81007
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This institutes a HistoryTracker mechanism that will allow the
FalsingManager to take recent activity into account when classifying
gestures. The HistoryTracker returns a penalty that is pased to each
classifier for them to use when making decisions.
As of right now, no classifier actually uses the history penalty when
making a classification decision.
The HistoryTracker keeps records of the last 3 seconds of falsing
results, applying a decaying penalty to the results over time. It also
assigns a confidence to its results, based on how consistent they are.
Each individual classifier is also able to return a confidence score
now, such that, if they are not confident in their decision, they
don't throw off the score that gets stored in the HistoryTracker.
Confidences currently returned by the classifiers are somewhat
arbitrary and should be adjusted in future CLs.
Everything is currently open for adjustment:
- The length of history
- The decay function applied to history
- The method by which our confidence in the history is calculated
- The confidence returned by each individual classfier
- How the history penalty is taken into account by each classifier
- ... and probably more.
Bug: 172655679
Test: atest SystemUITests
Change-Id: Ief2a5144a0cda659c7d78ed6e489231d481b6872
|
|
|
Sets up some more dagger injection and removes all the
"new Classifier(...)" from its constructor, allowing them
to come from outside.
Bug: 172655679
Test: atest SystemUITests
Change-Id: I916de7d5ee60ec2183c6401ec222867fc5eff823
|
