What about when the measuring scale is on the order of 20 to 30 cm rather than the Earth.
Even worse. Accelerometers are very bad for determining coordinates. Accelerometers return acceleration, and position is a second integral of acceleration. You have double integration errors. That's where the significant part of the error comes from, not the accelerometer itself.
Do they still get wildly inaccurate after awhile if you don't calibrate? By "accurate" I mean angular accuracy.
Yes, they are. Even if you have absolutely accurate accelerometer that never drifts, your position will drift because of numerical errors due to double integration.
Accelerometers are the worst method for this task.
That is not how an accelerometer is used for this. Attempting to determine a position from an accelerometer data is a futile effort unless working on millisecond time scales and having an external reference against which you can correct the rapidly accumulating error. **
An accelerometer is going to be used to determine the pitch and roll angles relative to gravity (the "down" vector). It won't drift and such solution is perfectly usable if you filter out the noise and the object is not accelerating or turning rapidly (the forces due to that will cause errors). On the other hand, an accelerometer won't give you the yaw (the rotation around the vertical axis). To get the yaw you need a magnetometer and use it as a magnetic compass (after compensating for pitch and roll which you get from the accelerometer).
Now if the object is moving, then you need a sensor that is immune to the forces due to turning and accelerations - and that is a gyro. Which will give you orientation no problem but it drifts. So the strategy is to use the data from the gyro most of the time but detect when the object is not accelerating/turning based on the sensor data and then take a fix from the accelerometer/magnetometer to zero out the gyro drift. And that's basically what an IMU does.
Or you mix the data proportionally based on how reliable they are - if accelerating fast, the accelerometer data quality is poor, so you use mostly gyro and only a little of the accelerometer/magnetometer. If not accelerating/turning, you give it higher confidence. This is, in essence, how a sensor fusion using something like a Kalman filter works.
(** That is very much how the HTC Vive/Valve Lighthouse laser-based virtual reality tracking system works - they use an accelerometer to estimate the distance traveled on very short timescales before the error gets out of hand - essentially dead-reckoning. Then whenever the laser beam from the base station hits the sensors on the tracked object they get an absolute position/orientation fix and use that to correct the estimated position. They use this strategy because it permits a very high update rate which is important for their application case. A typical accelerometer can do 1000+ readings/second whereas the laser is unable to scan that fast - it is mechanically swept about 100 times/second. But this is a rather atypical application.)