In a word… vaguely. The question of asteroid shape (even “hull,” or convexities only) is well-posed but still requires a significant amount of assumption to get “from here to there.” Hence, some asteroid models turn out to be pretty good in hindsight, and some turn out to be horrendous:
Lowry, S. et al. Dec. 2012 “The nucleus of Comet 67P/Churyumov-Gerasimenko. A new shape model and thermophysical analysis“ A & A 548, 12
Harris, A., Warner, B. 2018 "Asteroid lightcurves: can't tell a contact binary from a brick" AAS DPS Meeting, #414.03
In the immediate sense, taking a lightcurve gets you a rotation period… maybe. A non-sphere changes its reflection of sunlight twice per period, or perhaps four times if it’s aspheric but in a symmetric way (ambiguous Fourier solution). Hence it’s fairly easy to speak of a body’s aspect ratio (long dimension versus short dimension). But even that assumes no albedo features (bright or dark areas, messing with the lightcurve). Assuming flat reflectivity, this shape model is some egg shape, or possibly tangerine: aspheric in one dimension only.
On a longer timescale, space is inertial and the body’s rotation state is stable. (…Except when it isn’t: chaotic rotation or ‘tumbling.’) Over one orbit, the rotation axis will make one pivot about the sky. This additional period (plus the viewer’s state, on Earth) then gives a second curve, superimposed on the first. Deconvolving this second function gives some indication of the pole orientation… unless it doesn’t. Pole solutions often give a second (…first?) eigenvalue. The process continues, deriving a second-order aspect ratio which (may) give a second aspect ratio, for three lengths: the body’s a:b:c dimensions, expressed as ratios. This shape model then goes from egg, to mango or sunflower seed or similar.
Of course, nothing says the first or second lightcurves will be true sinusoids. Deviations from smooth sine waves are deviations from ellipsoidal profiles… except when they’re not. The higher-term irregularities may be local features like craters or such… or again, albedo differences.
And this is all glossing over instrumental effects and irreproducibilities- two observer groups will have two optical trains, with two profiles of detector noise and atmospheric ripple. Ultimately, we’d like a net lightcurve patched from an ensemble of corroborating results. The shape model, too, may have different groups doing their own deconvolutions. If the shape is reproducible enough between reviewers/follow-up work, we feel satisfied… until, of course, that rare Rosetta moment pulls the rug out. Fortunately, asteroids do have some gravity, and an asteroid big enough to draw study is usually big enough to not have reentrant forms like Churyumov-Gerasimenko.
If we’re lucky there will be a good, radar-worthy apparition while Goldstone, etc. is open. Then a range-dopplergram gives a second model, for comparison. Alternately, an occultation campaign may be able to put three or more chords across the body profile. If nontrivially placed, multiple chords constrain a profile to certain (convex) shapes, not others.