A cryptographic commitment scheme is a way of committing to a statement without revealing the statement itself or reveal it at a later time.
Or as Wikipedia explains it:
A way to visualize a commitment scheme is to think of a sender as putting a message in a locked box, and giving the box to a receiver. The message in the box is hidden from the receiver, who cannot open the lock himself. Since the receiver has the box, the message inside cannot be changedâ€”merely revealed if the sender chooses to give him the key at some later time.
In general a commitment scheme is executed in 3 phases:
 setup  the prover, the verifier (and sometimes a trusted thirdparty for secret parameters) agree on common system's parameters
 commit  the prover generates some random value r, commits to message m and calculates and sends the commitment c to the verifier

verify  also called decommit or open
 the prover reveals the message m and opening value r
 the verifier takes c, m and r and returns true/false whether the verification succeeds or not
Also, commitment schemes have two main properties that we need to consider:

hiding or confidentiality  verifier's capability to learn any information, given only the commitment c:
 perfectly hiding (or information theoretically hiding)  even an unbounded (with infinite computation power) adversary cannot reveal the committed message m
 computationally hiding  with infinite computation power an adversary can solve a hard problem (aka trapdoor function: prime factorization, discrete logarithm, etc) and find the committed message m

binding  prover's capability to find another message m' and open the commitment c without revealing the committed message m:
 perfectly binding  there is absolutely no way for an unbounded adversary to find another message m'
 computationally binding  an unbounded adversary can solve a hard problem and find another message m'
Let's take a simple example of rolling a dice where Peggy (the prover) commits to a value and Victor (the verifier) checks whether she wins of not.
A. Hashbased commitments
These are the simplest forms of commitment that have been around since forever.
1. Naive implementation
In setup phase the involved parties agree to use the 'SHA1' hashing:
import random
from hashlib import sha1
During the commit phase, Peggy 'calls' her bet m which will be kept secret for now, seals the bet and releases the commitment value c:
m = 6
c = sha1(bytes(m)).hexdigest()
print(c)
7722745105e9e02e8f1aaf17f7b3aac5c56cd805
Before the verify phase the following events happen:
 Victor rolls the dice, let's assume it shows a 6
 Peggy reveals the secret message m
Then, during verify phase Victor hashes the dice outcome and checks against commitment value c. If they are equal then Peggy won.
o = 6
check = sha1(bytes(o)).hexdigest()
print("Peggy won!") if c == check else "Try again!"
Peggy won!
Note: We can easily notice that naive hashbased commitment is both computationally hiding and binding becasue an adversary with unlimited computation power can execute a preimage attack (by brute forcing or other algorithms) and:
 as a verifier/receiver  find the committed message m (the hiding property)
 as a prover/sender  find another message m' to open the commitment (the binding property)
2. Salted implementation
So far so good but there is a problem with the above implementation: since all possible dice outputs are known in advance, Victor (or thirdparty) can precompute the hashes for all possible outcomes (known as Rainbow table attack) and bet against Peggy.
We can do better, instead of hashing only the committed message itself, we can concatenate with another random value called salt (or blinding factor) which will also be kept secret.
m = 6
s = random.randint(1, 9999)
c = sha1(bytes(m + s)).hexdigest()
print(c)
9a8e3678280e57c9b9f9d66f771a8f896f8517fe
This time, before the verify phase, Peggy releases both message value s and random value (or opening value) s.
o = 6
check = sha1(bytes(m + s)).hexdigest()
print("Peggy won!") if c == check else "Try again!"
Peggy won!
Note: Our salted hashbased commitment is also computationally hiding and binding because an unbounded adversary can find the concatenated value m + s then find the committed message m, also a prover/sender adversary can find another message m' to open the commitment.
B. Cryptosystembased commitments
It is known that any encryption scheme can be used as a commitment scheme, in our case we will look at noninteractive commitments, where the verifier/receiver does not send any information back to prover/sender/committer.
1. ElGamal commitment
During setup phase we agree on ElGamal's parameters. The p and g parameters are explained in ElGamal blog while h is another randomly chosen generator of prime order q.
from random import randint
p = 11
q = 7
g = 3
s = randint(1, q1)
h = g**s % p
print("h:", h, ", group:", [h**i % q for i in range(1, q)])
h: 5, group: [5, 4, 6, 2, 3, 1]
During commit phase, Peggy bets on message m, generates random value r and releases commitment tuple (c0, c1). Also, notice the blinding factor h**r that is thrown into the c1 mix.
m = 6
r = randint(1, q1)
c0 = ((g**r)) % p
c1 = ((g**m) * (h**r)) % p
print((c0, c1))
(3, 4)
Then roll the dice, Peggy reveals the message m and the opening value r and Victor checks the commitment, if equal, Peggy wins.
check = (g**m * h**r) % p
print("Peggy won!") if check == c1 else "Try again!"
Peggy won!
Note: maybe this is not so obvious but ElGamal commitment is computational hiding and perfectly binding because an unbounded adversary can try all possible values for r and solve the discrete logarithm (dlog) behind c0 commitment, then using r solve the dlog behind c1 commitment and figure out m.
2. Pedersen commitment
Pedersen commitment is similar to ElGamal above with a few differences that we will see below:
from random import randint
p = 11
q = 7
g = 3
s = randint(1, q1)
h = g**s % p
print(h)
5
Peggy commits on message m, generates random r and releases commitment c:
m = 6
r = randint(1, q1)
c = ((g**m) * (h**r)) % p
print(c)
4
At a later time the she reveals (m, r) tuple to Victor to verify the commitment.
check = (g**m * h**r) % p
print("Peggy won!") if check == c else "Try again!"
Peggy won!
Note: the Pedersen commitment is perfectly hiding and computationally binding because:
 as a verifier/receiver  has no way to uniquely identify the committed message m, there is an infinite number of (m, r) pairs that satisfy the commitment equation
 as a prover/sender  with infinite computation power can solve the dlog and find another message m'
So far, we can see that Pedersen commitment provides better security (perfectly hiding) than hashbased commitments but it has another property that makes it special.
2.1 Pedersen commitment on Elliptic Curve
Encryptionbased commitments are additively homomorphic which means that we can add the commitments together (while preserving the security and underlying properties), without actually opening the commitments.
Also, Pedersen commitment can be implemented with Elliptic curve where p, F and E parameters are explained in Elliptic curve blog post.
During the setup phase a random secret parameter s is used to create another elliptic curve generator H, the same as parameter h in exponential implementation, the difference is that H is a point on elliptic curve, not a scalar.
p = 103
F = FiniteField(p)
E = EllipticCurve(F, [0, 7])
print(E)
G = E([97, 10])
print(G)
s = randint(2, p)
H = s * G
print(H)
Elliptic Curve defined by y^2 = x^3 + 7 over Finite Field of size 103 (97 : 10 : 1) (33 : 10 : 1)
During the commit phase, generate random value r, commit to message m and release the commitment C, which is also a point on elliptic curve.
m = 2
r = randint(2, p)
C = m * H + r * G
print(C)
(53 : 7 : 1)
But this time we are not going to open the commitment, only leverage the additive homomorphic property, let's say to hide the amounts involved in a transaction, as shown in our final example below.
Peggy has 7 XMR on two different addresses 2 + 5 and wants to send 6 XMR to Victor and 1 XMR change back to her address but she also wants to keep it private so noone will be able to see what amounts are being transferred.
Peggy creates and releases the 4 commitments for each amount that is involved in transaction:
r = randint(2, p)
CI1 = 2 * H + r * G
CI2 = 5 * H + r * G
CO = 6 * H + r * G
CC = 1 * H + r * G
So far so good, now Victor has to verify if inputs are equal to outputs, if equal then transaction is valid and everybody is happy!
print "YOU ARE A CRYPTOSTAR!" if CI1 + CI2 == CO + CC else "YOU SUCK!"
YOU ARE A CRYPTOSTAR!
3. Switch commitments
Let's recap:
 hashbased commitments  provide both computationally hiding and binding
 ElGamal  is computationally hiding and perfectly binding
 Pedersen  is perfectly hiding and computationally binding
Alrightâ€¦.
 What about a commitment scheme that is both perfectly hiding and binding?
 I am afraid you won't like the answer because hiding and biding properties are mutually exclusive, you can only have one or the other, not both
 What the heck are switch commitments then?
 It is just a combination of ElGamal and Pedersen commitments with the ability to toggle one or the other on the fly in case shit hits the fan and quantum computers break encryption schemes over night
References
 https://en.wikipedia.org/wiki/Commitment_scheme
 https://crypto.stackexchange.com/questions/9704/whyisthepedersencommitmentcomputationallybinding?rq=1
 https://crypto.stackexchange.com/questions/11923/differencebetweenpedersencommitmentandcommitmentbasedonelgamal
 https://electriccoin.co/blog/snarkexplain/
 https://crypto.stackexchange.com/questions/60279/whatisthereasonofusingpedersencommitmentschemeoverhmac
 https://monero.stackexchange.com/questions/11213/whatistheimpactofhavingperfectlyhiddenbutonlycomputationallyboundcom
 https://github.com/mimblewimble/grin/blob/master/doc/switch_commitment.md