Ethernet started out as a protocol for coaxial shared medium networks. In the event of a collision the voltage levels on the cable would be wrong and the transceivers would detect this and tell the MACs. Any data received at this point would be garbage.
The receiving MAC would ignore any data coming in, and possibly increment a collision counter. The MAC has no reason to pass the garbage coming in from the PHY up to the host computer. The sending MACs would also see the collision, they would keep transmitting for the minimum frame time to ensure the collision was seen throughout the network. They would then wait for a random backoff before trying to send the frame again, hopefully the random backoff times used by the two sending MACs would be sufficiently different that one of them would succesfully send it's frame while the other would detect the line as busy and wait it's turn.
Now enter repeaters (aka hubs). A repeater has a number of PHYs but no MACs. When a repeater detects a collision it outputs a "Jam signal". This ensures that the collision is seen across the whole network.
So what about twisted pair? All common variants of twisted pair Ethernet are full duplex at the physical layer but for backwards compatibility and to support the use of hubs most of them have a "half duplex" mode. In half duplex mode simultaneous activation of transmitter and receiver is treated as a collision.
Putting that all together lets imagine you had a device plugged in to one of the spare ports on your hub that could capture and analyse low level Ethernet signals, what would it see?
You would see the start of a frame as normal. Then when the collision happened you would see a Jam signal from the hub. Once the Jam signal was over you would then see a period of idle line until one of the sender's random back-off timers expires and a retransmission begins.